2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/cpuidle.h>
27 #include <linux/slab.h>
28 #include <linux/profile.h>
29 #include <linux/interrupt.h>
30 #include <linux/mempolicy.h>
31 #include <linux/migrate.h>
32 #include <linux/task_work.h>
33 #include <linux/module.h>
35 #include <trace/events/sched.h>
42 * Targeted preemption latency for CPU-bound tasks:
43 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
45 * NOTE: this latency value is not the same as the concept of
46 * 'timeslice length' - timeslices in CFS are of variable length
47 * and have no persistent notion like in traditional, time-slice
48 * based scheduling concepts.
50 * (to see the precise effective timeslice length of your workload,
51 * run vmstat and monitor the context-switches (cs) field)
53 unsigned int sysctl_sched_latency = 6000000ULL;
54 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
56 unsigned int sysctl_sched_is_big_little = 0;
57 unsigned int sysctl_sched_sync_hint_enable = 1;
58 unsigned int sysctl_sched_initial_task_util = 0;
59 unsigned int sysctl_sched_cstate_aware = 1;
61 #ifdef CONFIG_SCHED_WALT
62 unsigned int sysctl_sched_use_walt_cpu_util = 1;
63 unsigned int sysctl_sched_use_walt_task_util = 1;
64 __read_mostly unsigned int sysctl_sched_walt_cpu_high_irqload =
68 * The initial- and re-scaling of tunables is configurable
69 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
72 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
73 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
74 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
76 enum sched_tunable_scaling sysctl_sched_tunable_scaling
77 = SCHED_TUNABLESCALING_LOG;
80 * Minimal preemption granularity for CPU-bound tasks:
81 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 unsigned int sysctl_sched_min_granularity = 750000ULL;
84 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
87 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
89 static unsigned int sched_nr_latency = 8;
92 * After fork, child runs first. If set to 0 (default) then
93 * parent will (try to) run first.
95 unsigned int sysctl_sched_child_runs_first __read_mostly;
98 * SCHED_OTHER wake-up granularity.
99 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
101 * This option delays the preemption effects of decoupled workloads
102 * and reduces their over-scheduling. Synchronous workloads will still
103 * have immediate wakeup/sleep latencies.
105 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
106 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
108 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
111 * The exponential sliding window over which load is averaged for shares
115 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
117 #ifdef CONFIG_CFS_BANDWIDTH
119 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
120 * each time a cfs_rq requests quota.
122 * Note: in the case that the slice exceeds the runtime remaining (either due
123 * to consumption or the quota being specified to be smaller than the slice)
124 * we will always only issue the remaining available time.
126 * default: 5 msec, units: microseconds
128 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
132 * The margin used when comparing utilization with CPU capacity:
133 * util * margin < capacity * 1024
135 unsigned int capacity_margin = 1280; /* ~20% */
137 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
143 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
149 static inline void update_load_set(struct load_weight *lw, unsigned long w)
156 * Increase the granularity value when there are more CPUs,
157 * because with more CPUs the 'effective latency' as visible
158 * to users decreases. But the relationship is not linear,
159 * so pick a second-best guess by going with the log2 of the
162 * This idea comes from the SD scheduler of Con Kolivas:
164 static unsigned int get_update_sysctl_factor(void)
166 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
169 switch (sysctl_sched_tunable_scaling) {
170 case SCHED_TUNABLESCALING_NONE:
173 case SCHED_TUNABLESCALING_LINEAR:
176 case SCHED_TUNABLESCALING_LOG:
178 factor = 1 + ilog2(cpus);
185 static void update_sysctl(void)
187 unsigned int factor = get_update_sysctl_factor();
189 #define SET_SYSCTL(name) \
190 (sysctl_##name = (factor) * normalized_sysctl_##name)
191 SET_SYSCTL(sched_min_granularity);
192 SET_SYSCTL(sched_latency);
193 SET_SYSCTL(sched_wakeup_granularity);
197 void sched_init_granularity(void)
202 #define WMULT_CONST (~0U)
203 #define WMULT_SHIFT 32
205 static void __update_inv_weight(struct load_weight *lw)
209 if (likely(lw->inv_weight))
212 w = scale_load_down(lw->weight);
214 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
216 else if (unlikely(!w))
217 lw->inv_weight = WMULT_CONST;
219 lw->inv_weight = WMULT_CONST / w;
223 * delta_exec * weight / lw.weight
225 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
227 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
228 * we're guaranteed shift stays positive because inv_weight is guaranteed to
229 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
231 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
232 * weight/lw.weight <= 1, and therefore our shift will also be positive.
234 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
236 u64 fact = scale_load_down(weight);
237 int shift = WMULT_SHIFT;
239 __update_inv_weight(lw);
241 if (unlikely(fact >> 32)) {
248 /* hint to use a 32x32->64 mul */
249 fact = (u64)(u32)fact * lw->inv_weight;
256 return mul_u64_u32_shr(delta_exec, fact, shift);
260 const struct sched_class fair_sched_class;
262 /**************************************************************
263 * CFS operations on generic schedulable entities:
266 #ifdef CONFIG_FAIR_GROUP_SCHED
268 /* cpu runqueue to which this cfs_rq is attached */
269 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
274 /* An entity is a task if it doesn't "own" a runqueue */
275 #define entity_is_task(se) (!se->my_q)
277 static inline struct task_struct *task_of(struct sched_entity *se)
279 #ifdef CONFIG_SCHED_DEBUG
280 WARN_ON_ONCE(!entity_is_task(se));
282 return container_of(se, struct task_struct, se);
285 /* Walk up scheduling entities hierarchy */
286 #define for_each_sched_entity(se) \
287 for (; se; se = se->parent)
289 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
294 /* runqueue on which this entity is (to be) queued */
295 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
300 /* runqueue "owned" by this group */
301 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
306 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
308 if (!cfs_rq->on_list) {
310 * Ensure we either appear before our parent (if already
311 * enqueued) or force our parent to appear after us when it is
312 * enqueued. The fact that we always enqueue bottom-up
313 * reduces this to two cases.
315 if (cfs_rq->tg->parent &&
316 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
317 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
318 &rq_of(cfs_rq)->leaf_cfs_rq_list);
320 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
321 &rq_of(cfs_rq)->leaf_cfs_rq_list);
328 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
330 if (cfs_rq->on_list) {
331 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
336 /* Iterate thr' all leaf cfs_rq's on a runqueue */
337 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
338 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
340 /* Do the two (enqueued) entities belong to the same group ? */
341 static inline struct cfs_rq *
342 is_same_group(struct sched_entity *se, struct sched_entity *pse)
344 if (se->cfs_rq == pse->cfs_rq)
350 static inline struct sched_entity *parent_entity(struct sched_entity *se)
356 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
358 int se_depth, pse_depth;
361 * preemption test can be made between sibling entities who are in the
362 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
363 * both tasks until we find their ancestors who are siblings of common
367 /* First walk up until both entities are at same depth */
368 se_depth = (*se)->depth;
369 pse_depth = (*pse)->depth;
371 while (se_depth > pse_depth) {
373 *se = parent_entity(*se);
376 while (pse_depth > se_depth) {
378 *pse = parent_entity(*pse);
381 while (!is_same_group(*se, *pse)) {
382 *se = parent_entity(*se);
383 *pse = parent_entity(*pse);
387 #else /* !CONFIG_FAIR_GROUP_SCHED */
389 static inline struct task_struct *task_of(struct sched_entity *se)
391 return container_of(se, struct task_struct, se);
394 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
396 return container_of(cfs_rq, struct rq, cfs);
399 #define entity_is_task(se) 1
401 #define for_each_sched_entity(se) \
402 for (; se; se = NULL)
404 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
406 return &task_rq(p)->cfs;
409 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
411 struct task_struct *p = task_of(se);
412 struct rq *rq = task_rq(p);
417 /* runqueue "owned" by this group */
418 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
423 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
427 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
431 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
432 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
434 static inline struct sched_entity *parent_entity(struct sched_entity *se)
440 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
444 #endif /* CONFIG_FAIR_GROUP_SCHED */
446 static __always_inline
447 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
449 /**************************************************************
450 * Scheduling class tree data structure manipulation methods:
453 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
455 s64 delta = (s64)(vruntime - max_vruntime);
457 max_vruntime = vruntime;
462 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
464 s64 delta = (s64)(vruntime - min_vruntime);
466 min_vruntime = vruntime;
471 static inline int entity_before(struct sched_entity *a,
472 struct sched_entity *b)
474 return (s64)(a->vruntime - b->vruntime) < 0;
477 static void update_min_vruntime(struct cfs_rq *cfs_rq)
479 u64 vruntime = cfs_rq->min_vruntime;
482 vruntime = cfs_rq->curr->vruntime;
484 if (cfs_rq->rb_leftmost) {
485 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
490 vruntime = se->vruntime;
492 vruntime = min_vruntime(vruntime, se->vruntime);
495 /* ensure we never gain time by being placed backwards. */
496 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
499 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
504 * Enqueue an entity into the rb-tree:
506 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
508 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
509 struct rb_node *parent = NULL;
510 struct sched_entity *entry;
514 * Find the right place in the rbtree:
518 entry = rb_entry(parent, struct sched_entity, run_node);
520 * We dont care about collisions. Nodes with
521 * the same key stay together.
523 if (entity_before(se, entry)) {
524 link = &parent->rb_left;
526 link = &parent->rb_right;
532 * Maintain a cache of leftmost tree entries (it is frequently
536 cfs_rq->rb_leftmost = &se->run_node;
538 rb_link_node(&se->run_node, parent, link);
539 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
542 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
544 if (cfs_rq->rb_leftmost == &se->run_node) {
545 struct rb_node *next_node;
547 next_node = rb_next(&se->run_node);
548 cfs_rq->rb_leftmost = next_node;
551 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
554 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
556 struct rb_node *left = cfs_rq->rb_leftmost;
561 return rb_entry(left, struct sched_entity, run_node);
564 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
566 struct rb_node *next = rb_next(&se->run_node);
571 return rb_entry(next, struct sched_entity, run_node);
574 #ifdef CONFIG_SCHED_DEBUG
575 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
577 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
582 return rb_entry(last, struct sched_entity, run_node);
585 /**************************************************************
586 * Scheduling class statistics methods:
589 int sched_proc_update_handler(struct ctl_table *table, int write,
590 void __user *buffer, size_t *lenp,
593 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
594 unsigned int factor = get_update_sysctl_factor();
599 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
600 sysctl_sched_min_granularity);
602 #define WRT_SYSCTL(name) \
603 (normalized_sysctl_##name = sysctl_##name / (factor))
604 WRT_SYSCTL(sched_min_granularity);
605 WRT_SYSCTL(sched_latency);
606 WRT_SYSCTL(sched_wakeup_granularity);
616 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 if (unlikely(nr_running > sched_nr_latency))
635 return nr_running * sysctl_sched_min_granularity;
637 return sysctl_sched_latency;
641 * We calculate the wall-time slice from the period by taking a part
642 * proportional to the weight.
646 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
648 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
650 for_each_sched_entity(se) {
651 struct load_weight *load;
652 struct load_weight lw;
654 cfs_rq = cfs_rq_of(se);
655 load = &cfs_rq->load;
657 if (unlikely(!se->on_rq)) {
660 update_load_add(&lw, se->load.weight);
663 slice = __calc_delta(slice, se->load.weight, load);
669 * We calculate the vruntime slice of a to-be-inserted task.
673 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
675 return calc_delta_fair(sched_slice(cfs_rq, se), se);
679 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
680 static unsigned long task_h_load(struct task_struct *p);
683 * We choose a half-life close to 1 scheduling period.
684 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
685 * dependent on this value.
687 #define LOAD_AVG_PERIOD 32
688 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
689 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
691 /* Give new sched_entity start runnable values to heavy its load in infant time */
692 void init_entity_runnable_average(struct sched_entity *se)
694 struct sched_avg *sa = &se->avg;
696 sa->last_update_time = 0;
698 * sched_avg's period_contrib should be strictly less then 1024, so
699 * we give it 1023 to make sure it is almost a period (1024us), and
700 * will definitely be update (after enqueue).
702 sa->period_contrib = 1023;
703 sa->load_avg = scale_load_down(se->load.weight);
704 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
705 sa->util_avg = sched_freq() ?
706 sysctl_sched_initial_task_util :
707 scale_load_down(SCHED_LOAD_SCALE);
708 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
709 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
713 void init_entity_runnable_average(struct sched_entity *se)
719 * Update the current task's runtime statistics.
721 static void update_curr(struct cfs_rq *cfs_rq)
723 struct sched_entity *curr = cfs_rq->curr;
724 u64 now = rq_clock_task(rq_of(cfs_rq));
730 delta_exec = now - curr->exec_start;
731 if (unlikely((s64)delta_exec <= 0))
734 curr->exec_start = now;
736 schedstat_set(curr->statistics.exec_max,
737 max(delta_exec, curr->statistics.exec_max));
739 curr->sum_exec_runtime += delta_exec;
740 schedstat_add(cfs_rq, exec_clock, delta_exec);
742 curr->vruntime += calc_delta_fair(delta_exec, curr);
743 update_min_vruntime(cfs_rq);
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
756 static void update_curr_fair(struct rq *rq)
758 update_curr(cfs_rq_of(&rq->curr->se));
762 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
764 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
768 * Task is being enqueued - update stats:
770 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 * Are we enqueueing a waiting task? (for current tasks
774 * a dequeue/enqueue event is a NOP)
776 if (se != cfs_rq->curr)
777 update_stats_wait_start(cfs_rq, se);
781 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
783 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
785 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
786 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
787 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
788 #ifdef CONFIG_SCHEDSTATS
789 if (entity_is_task(se)) {
790 trace_sched_stat_wait(task_of(se),
791 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
794 schedstat_set(se->statistics.wait_start, 0);
798 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
801 * Mark the end of the wait period if dequeueing a
804 if (se != cfs_rq->curr)
805 update_stats_wait_end(cfs_rq, se);
809 * We are picking a new current task - update its stats:
812 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
815 * We are starting a new run period:
817 se->exec_start = rq_clock_task(rq_of(cfs_rq));
820 /**************************************************
821 * Scheduling class queueing methods:
824 #ifdef CONFIG_NUMA_BALANCING
826 * Approximate time to scan a full NUMA task in ms. The task scan period is
827 * calculated based on the tasks virtual memory size and
828 * numa_balancing_scan_size.
830 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
831 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
833 /* Portion of address space to scan in MB */
834 unsigned int sysctl_numa_balancing_scan_size = 256;
836 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
837 unsigned int sysctl_numa_balancing_scan_delay = 1000;
839 static unsigned int task_nr_scan_windows(struct task_struct *p)
841 unsigned long rss = 0;
842 unsigned long nr_scan_pages;
845 * Calculations based on RSS as non-present and empty pages are skipped
846 * by the PTE scanner and NUMA hinting faults should be trapped based
849 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
850 rss = get_mm_rss(p->mm);
854 rss = round_up(rss, nr_scan_pages);
855 return rss / nr_scan_pages;
858 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
859 #define MAX_SCAN_WINDOW 2560
861 static unsigned int task_scan_min(struct task_struct *p)
863 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
864 unsigned int scan, floor;
865 unsigned int windows = 1;
867 if (scan_size < MAX_SCAN_WINDOW)
868 windows = MAX_SCAN_WINDOW / scan_size;
869 floor = 1000 / windows;
871 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
872 return max_t(unsigned int, floor, scan);
875 static unsigned int task_scan_max(struct task_struct *p)
877 unsigned int smin = task_scan_min(p);
880 /* Watch for min being lower than max due to floor calculations */
881 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
882 return max(smin, smax);
885 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
887 rq->nr_numa_running += (p->numa_preferred_nid != -1);
888 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
891 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
893 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
894 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
900 spinlock_t lock; /* nr_tasks, tasks */
905 nodemask_t active_nodes;
906 unsigned long total_faults;
908 * Faults_cpu is used to decide whether memory should move
909 * towards the CPU. As a consequence, these stats are weighted
910 * more by CPU use than by memory faults.
912 unsigned long *faults_cpu;
913 unsigned long faults[0];
916 /* Shared or private faults. */
917 #define NR_NUMA_HINT_FAULT_TYPES 2
919 /* Memory and CPU locality */
920 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
922 /* Averaged statistics, and temporary buffers. */
923 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
925 pid_t task_numa_group_id(struct task_struct *p)
927 return p->numa_group ? p->numa_group->gid : 0;
931 * The averaged statistics, shared & private, memory & cpu,
932 * occupy the first half of the array. The second half of the
933 * array is for current counters, which are averaged into the
934 * first set by task_numa_placement.
936 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
938 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
941 static inline unsigned long task_faults(struct task_struct *p, int nid)
946 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
947 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
950 static inline unsigned long group_faults(struct task_struct *p, int nid)
955 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
956 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
959 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
961 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
962 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
965 /* Handle placement on systems where not all nodes are directly connected. */
966 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
967 int maxdist, bool task)
969 unsigned long score = 0;
973 * All nodes are directly connected, and the same distance
974 * from each other. No need for fancy placement algorithms.
976 if (sched_numa_topology_type == NUMA_DIRECT)
980 * This code is called for each node, introducing N^2 complexity,
981 * which should be ok given the number of nodes rarely exceeds 8.
983 for_each_online_node(node) {
984 unsigned long faults;
985 int dist = node_distance(nid, node);
988 * The furthest away nodes in the system are not interesting
989 * for placement; nid was already counted.
991 if (dist == sched_max_numa_distance || node == nid)
995 * On systems with a backplane NUMA topology, compare groups
996 * of nodes, and move tasks towards the group with the most
997 * memory accesses. When comparing two nodes at distance
998 * "hoplimit", only nodes closer by than "hoplimit" are part
999 * of each group. Skip other nodes.
1001 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1005 /* Add up the faults from nearby nodes. */
1007 faults = task_faults(p, node);
1009 faults = group_faults(p, node);
1012 * On systems with a glueless mesh NUMA topology, there are
1013 * no fixed "groups of nodes". Instead, nodes that are not
1014 * directly connected bounce traffic through intermediate
1015 * nodes; a numa_group can occupy any set of nodes.
1016 * The further away a node is, the less the faults count.
1017 * This seems to result in good task placement.
1019 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1020 faults *= (sched_max_numa_distance - dist);
1021 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1031 * These return the fraction of accesses done by a particular task, or
1032 * task group, on a particular numa node. The group weight is given a
1033 * larger multiplier, in order to group tasks together that are almost
1034 * evenly spread out between numa nodes.
1036 static inline unsigned long task_weight(struct task_struct *p, int nid,
1039 unsigned long faults, total_faults;
1041 if (!p->numa_faults)
1044 total_faults = p->total_numa_faults;
1049 faults = task_faults(p, nid);
1050 faults += score_nearby_nodes(p, nid, dist, true);
1052 return 1000 * faults / total_faults;
1055 static inline unsigned long group_weight(struct task_struct *p, int nid,
1058 unsigned long faults, total_faults;
1063 total_faults = p->numa_group->total_faults;
1068 faults = group_faults(p, nid);
1069 faults += score_nearby_nodes(p, nid, dist, false);
1071 return 1000 * faults / total_faults;
1074 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1075 int src_nid, int dst_cpu)
1077 struct numa_group *ng = p->numa_group;
1078 int dst_nid = cpu_to_node(dst_cpu);
1079 int last_cpupid, this_cpupid;
1081 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1084 * Multi-stage node selection is used in conjunction with a periodic
1085 * migration fault to build a temporal task<->page relation. By using
1086 * a two-stage filter we remove short/unlikely relations.
1088 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1089 * a task's usage of a particular page (n_p) per total usage of this
1090 * page (n_t) (in a given time-span) to a probability.
1092 * Our periodic faults will sample this probability and getting the
1093 * same result twice in a row, given these samples are fully
1094 * independent, is then given by P(n)^2, provided our sample period
1095 * is sufficiently short compared to the usage pattern.
1097 * This quadric squishes small probabilities, making it less likely we
1098 * act on an unlikely task<->page relation.
1100 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1101 if (!cpupid_pid_unset(last_cpupid) &&
1102 cpupid_to_nid(last_cpupid) != dst_nid)
1105 /* Always allow migrate on private faults */
1106 if (cpupid_match_pid(p, last_cpupid))
1109 /* A shared fault, but p->numa_group has not been set up yet. */
1114 * Do not migrate if the destination is not a node that
1115 * is actively used by this numa group.
1117 if (!node_isset(dst_nid, ng->active_nodes))
1121 * Source is a node that is not actively used by this
1122 * numa group, while the destination is. Migrate.
1124 if (!node_isset(src_nid, ng->active_nodes))
1128 * Both source and destination are nodes in active
1129 * use by this numa group. Maximize memory bandwidth
1130 * by migrating from more heavily used groups, to less
1131 * heavily used ones, spreading the load around.
1132 * Use a 1/4 hysteresis to avoid spurious page movement.
1134 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1137 static unsigned long weighted_cpuload(const int cpu);
1138 static unsigned long source_load(int cpu, int type);
1139 static unsigned long target_load(int cpu, int type);
1140 static unsigned long capacity_of(int cpu);
1141 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1143 /* Cached statistics for all CPUs within a node */
1145 unsigned long nr_running;
1148 /* Total compute capacity of CPUs on a node */
1149 unsigned long compute_capacity;
1151 /* Approximate capacity in terms of runnable tasks on a node */
1152 unsigned long task_capacity;
1153 int has_free_capacity;
1157 * XXX borrowed from update_sg_lb_stats
1159 static void update_numa_stats(struct numa_stats *ns, int nid)
1161 int smt, cpu, cpus = 0;
1162 unsigned long capacity;
1164 memset(ns, 0, sizeof(*ns));
1165 for_each_cpu(cpu, cpumask_of_node(nid)) {
1166 struct rq *rq = cpu_rq(cpu);
1168 ns->nr_running += rq->nr_running;
1169 ns->load += weighted_cpuload(cpu);
1170 ns->compute_capacity += capacity_of(cpu);
1176 * If we raced with hotplug and there are no CPUs left in our mask
1177 * the @ns structure is NULL'ed and task_numa_compare() will
1178 * not find this node attractive.
1180 * We'll either bail at !has_free_capacity, or we'll detect a huge
1181 * imbalance and bail there.
1186 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1187 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1188 capacity = cpus / smt; /* cores */
1190 ns->task_capacity = min_t(unsigned, capacity,
1191 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1192 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1195 struct task_numa_env {
1196 struct task_struct *p;
1198 int src_cpu, src_nid;
1199 int dst_cpu, dst_nid;
1201 struct numa_stats src_stats, dst_stats;
1206 struct task_struct *best_task;
1211 static void task_numa_assign(struct task_numa_env *env,
1212 struct task_struct *p, long imp)
1215 put_task_struct(env->best_task);
1218 env->best_imp = imp;
1219 env->best_cpu = env->dst_cpu;
1222 static bool load_too_imbalanced(long src_load, long dst_load,
1223 struct task_numa_env *env)
1226 long orig_src_load, orig_dst_load;
1227 long src_capacity, dst_capacity;
1230 * The load is corrected for the CPU capacity available on each node.
1233 * ------------ vs ---------
1234 * src_capacity dst_capacity
1236 src_capacity = env->src_stats.compute_capacity;
1237 dst_capacity = env->dst_stats.compute_capacity;
1239 /* We care about the slope of the imbalance, not the direction. */
1240 if (dst_load < src_load)
1241 swap(dst_load, src_load);
1243 /* Is the difference below the threshold? */
1244 imb = dst_load * src_capacity * 100 -
1245 src_load * dst_capacity * env->imbalance_pct;
1250 * The imbalance is above the allowed threshold.
1251 * Compare it with the old imbalance.
1253 orig_src_load = env->src_stats.load;
1254 orig_dst_load = env->dst_stats.load;
1256 if (orig_dst_load < orig_src_load)
1257 swap(orig_dst_load, orig_src_load);
1259 old_imb = orig_dst_load * src_capacity * 100 -
1260 orig_src_load * dst_capacity * env->imbalance_pct;
1262 /* Would this change make things worse? */
1263 return (imb > old_imb);
1267 * This checks if the overall compute and NUMA accesses of the system would
1268 * be improved if the source tasks was migrated to the target dst_cpu taking
1269 * into account that it might be best if task running on the dst_cpu should
1270 * be exchanged with the source task
1272 static void task_numa_compare(struct task_numa_env *env,
1273 long taskimp, long groupimp)
1275 struct rq *src_rq = cpu_rq(env->src_cpu);
1276 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1277 struct task_struct *cur;
1278 long src_load, dst_load;
1280 long imp = env->p->numa_group ? groupimp : taskimp;
1282 int dist = env->dist;
1283 bool assigned = false;
1287 raw_spin_lock_irq(&dst_rq->lock);
1290 * No need to move the exiting task or idle task.
1292 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1296 * The task_struct must be protected here to protect the
1297 * p->numa_faults access in the task_weight since the
1298 * numa_faults could already be freed in the following path:
1299 * finish_task_switch()
1300 * --> put_task_struct()
1301 * --> __put_task_struct()
1302 * --> task_numa_free()
1304 get_task_struct(cur);
1307 raw_spin_unlock_irq(&dst_rq->lock);
1310 * Because we have preemption enabled we can get migrated around and
1311 * end try selecting ourselves (current == env->p) as a swap candidate.
1317 * "imp" is the fault differential for the source task between the
1318 * source and destination node. Calculate the total differential for
1319 * the source task and potential destination task. The more negative
1320 * the value is, the more rmeote accesses that would be expected to
1321 * be incurred if the tasks were swapped.
1324 /* Skip this swap candidate if cannot move to the source cpu */
1325 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1329 * If dst and source tasks are in the same NUMA group, or not
1330 * in any group then look only at task weights.
1332 if (cur->numa_group == env->p->numa_group) {
1333 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1334 task_weight(cur, env->dst_nid, dist);
1336 * Add some hysteresis to prevent swapping the
1337 * tasks within a group over tiny differences.
1339 if (cur->numa_group)
1343 * Compare the group weights. If a task is all by
1344 * itself (not part of a group), use the task weight
1347 if (cur->numa_group)
1348 imp += group_weight(cur, env->src_nid, dist) -
1349 group_weight(cur, env->dst_nid, dist);
1351 imp += task_weight(cur, env->src_nid, dist) -
1352 task_weight(cur, env->dst_nid, dist);
1356 if (imp <= env->best_imp && moveimp <= env->best_imp)
1360 /* Is there capacity at our destination? */
1361 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1362 !env->dst_stats.has_free_capacity)
1368 /* Balance doesn't matter much if we're running a task per cpu */
1369 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1370 dst_rq->nr_running == 1)
1374 * In the overloaded case, try and keep the load balanced.
1377 load = task_h_load(env->p);
1378 dst_load = env->dst_stats.load + load;
1379 src_load = env->src_stats.load - load;
1381 if (moveimp > imp && moveimp > env->best_imp) {
1383 * If the improvement from just moving env->p direction is
1384 * better than swapping tasks around, check if a move is
1385 * possible. Store a slightly smaller score than moveimp,
1386 * so an actually idle CPU will win.
1388 if (!load_too_imbalanced(src_load, dst_load, env)) {
1390 put_task_struct(cur);
1396 if (imp <= env->best_imp)
1400 load = task_h_load(cur);
1405 if (load_too_imbalanced(src_load, dst_load, env))
1409 * One idle CPU per node is evaluated for a task numa move.
1410 * Call select_idle_sibling to maybe find a better one.
1413 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1418 task_numa_assign(env, cur, imp);
1422 * The dst_rq->curr isn't assigned. The protection for task_struct is
1425 if (cur && !assigned)
1426 put_task_struct(cur);
1429 static void task_numa_find_cpu(struct task_numa_env *env,
1430 long taskimp, long groupimp)
1434 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1435 /* Skip this CPU if the source task cannot migrate */
1436 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1440 task_numa_compare(env, taskimp, groupimp);
1444 /* Only move tasks to a NUMA node less busy than the current node. */
1445 static bool numa_has_capacity(struct task_numa_env *env)
1447 struct numa_stats *src = &env->src_stats;
1448 struct numa_stats *dst = &env->dst_stats;
1450 if (src->has_free_capacity && !dst->has_free_capacity)
1454 * Only consider a task move if the source has a higher load
1455 * than the destination, corrected for CPU capacity on each node.
1457 * src->load dst->load
1458 * --------------------- vs ---------------------
1459 * src->compute_capacity dst->compute_capacity
1461 if (src->load * dst->compute_capacity * env->imbalance_pct >
1463 dst->load * src->compute_capacity * 100)
1469 static int task_numa_migrate(struct task_struct *p)
1471 struct task_numa_env env = {
1474 .src_cpu = task_cpu(p),
1475 .src_nid = task_node(p),
1477 .imbalance_pct = 112,
1483 struct sched_domain *sd;
1484 unsigned long taskweight, groupweight;
1486 long taskimp, groupimp;
1489 * Pick the lowest SD_NUMA domain, as that would have the smallest
1490 * imbalance and would be the first to start moving tasks about.
1492 * And we want to avoid any moving of tasks about, as that would create
1493 * random movement of tasks -- counter the numa conditions we're trying
1497 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1499 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1503 * Cpusets can break the scheduler domain tree into smaller
1504 * balance domains, some of which do not cross NUMA boundaries.
1505 * Tasks that are "trapped" in such domains cannot be migrated
1506 * elsewhere, so there is no point in (re)trying.
1508 if (unlikely(!sd)) {
1509 p->numa_preferred_nid = task_node(p);
1513 env.dst_nid = p->numa_preferred_nid;
1514 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1515 taskweight = task_weight(p, env.src_nid, dist);
1516 groupweight = group_weight(p, env.src_nid, dist);
1517 update_numa_stats(&env.src_stats, env.src_nid);
1518 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1519 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1520 update_numa_stats(&env.dst_stats, env.dst_nid);
1522 /* Try to find a spot on the preferred nid. */
1523 if (numa_has_capacity(&env))
1524 task_numa_find_cpu(&env, taskimp, groupimp);
1527 * Look at other nodes in these cases:
1528 * - there is no space available on the preferred_nid
1529 * - the task is part of a numa_group that is interleaved across
1530 * multiple NUMA nodes; in order to better consolidate the group,
1531 * we need to check other locations.
1533 if (env.best_cpu == -1 || (p->numa_group &&
1534 nodes_weight(p->numa_group->active_nodes) > 1)) {
1535 for_each_online_node(nid) {
1536 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1539 dist = node_distance(env.src_nid, env.dst_nid);
1540 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1542 taskweight = task_weight(p, env.src_nid, dist);
1543 groupweight = group_weight(p, env.src_nid, dist);
1546 /* Only consider nodes where both task and groups benefit */
1547 taskimp = task_weight(p, nid, dist) - taskweight;
1548 groupimp = group_weight(p, nid, dist) - groupweight;
1549 if (taskimp < 0 && groupimp < 0)
1554 update_numa_stats(&env.dst_stats, env.dst_nid);
1555 if (numa_has_capacity(&env))
1556 task_numa_find_cpu(&env, taskimp, groupimp);
1561 * If the task is part of a workload that spans multiple NUMA nodes,
1562 * and is migrating into one of the workload's active nodes, remember
1563 * this node as the task's preferred numa node, so the workload can
1565 * A task that migrated to a second choice node will be better off
1566 * trying for a better one later. Do not set the preferred node here.
1568 if (p->numa_group) {
1569 if (env.best_cpu == -1)
1574 if (node_isset(nid, p->numa_group->active_nodes))
1575 sched_setnuma(p, env.dst_nid);
1578 /* No better CPU than the current one was found. */
1579 if (env.best_cpu == -1)
1583 * Reset the scan period if the task is being rescheduled on an
1584 * alternative node to recheck if the tasks is now properly placed.
1586 p->numa_scan_period = task_scan_min(p);
1588 if (env.best_task == NULL) {
1589 ret = migrate_task_to(p, env.best_cpu);
1591 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1595 ret = migrate_swap(p, env.best_task);
1597 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1598 put_task_struct(env.best_task);
1602 /* Attempt to migrate a task to a CPU on the preferred node. */
1603 static void numa_migrate_preferred(struct task_struct *p)
1605 unsigned long interval = HZ;
1607 /* This task has no NUMA fault statistics yet */
1608 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1611 /* Periodically retry migrating the task to the preferred node */
1612 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1613 p->numa_migrate_retry = jiffies + interval;
1615 /* Success if task is already running on preferred CPU */
1616 if (task_node(p) == p->numa_preferred_nid)
1619 /* Otherwise, try migrate to a CPU on the preferred node */
1620 task_numa_migrate(p);
1624 * Find the nodes on which the workload is actively running. We do this by
1625 * tracking the nodes from which NUMA hinting faults are triggered. This can
1626 * be different from the set of nodes where the workload's memory is currently
1629 * The bitmask is used to make smarter decisions on when to do NUMA page
1630 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1631 * are added when they cause over 6/16 of the maximum number of faults, but
1632 * only removed when they drop below 3/16.
1634 static void update_numa_active_node_mask(struct numa_group *numa_group)
1636 unsigned long faults, max_faults = 0;
1639 for_each_online_node(nid) {
1640 faults = group_faults_cpu(numa_group, nid);
1641 if (faults > max_faults)
1642 max_faults = faults;
1645 for_each_online_node(nid) {
1646 faults = group_faults_cpu(numa_group, nid);
1647 if (!node_isset(nid, numa_group->active_nodes)) {
1648 if (faults > max_faults * 6 / 16)
1649 node_set(nid, numa_group->active_nodes);
1650 } else if (faults < max_faults * 3 / 16)
1651 node_clear(nid, numa_group->active_nodes);
1656 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1657 * increments. The more local the fault statistics are, the higher the scan
1658 * period will be for the next scan window. If local/(local+remote) ratio is
1659 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1660 * the scan period will decrease. Aim for 70% local accesses.
1662 #define NUMA_PERIOD_SLOTS 10
1663 #define NUMA_PERIOD_THRESHOLD 7
1666 * Increase the scan period (slow down scanning) if the majority of
1667 * our memory is already on our local node, or if the majority of
1668 * the page accesses are shared with other processes.
1669 * Otherwise, decrease the scan period.
1671 static void update_task_scan_period(struct task_struct *p,
1672 unsigned long shared, unsigned long private)
1674 unsigned int period_slot;
1678 unsigned long remote = p->numa_faults_locality[0];
1679 unsigned long local = p->numa_faults_locality[1];
1682 * If there were no record hinting faults then either the task is
1683 * completely idle or all activity is areas that are not of interest
1684 * to automatic numa balancing. Related to that, if there were failed
1685 * migration then it implies we are migrating too quickly or the local
1686 * node is overloaded. In either case, scan slower
1688 if (local + shared == 0 || p->numa_faults_locality[2]) {
1689 p->numa_scan_period = min(p->numa_scan_period_max,
1690 p->numa_scan_period << 1);
1692 p->mm->numa_next_scan = jiffies +
1693 msecs_to_jiffies(p->numa_scan_period);
1699 * Prepare to scale scan period relative to the current period.
1700 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1701 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1702 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1704 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1705 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1706 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1707 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1710 diff = slot * period_slot;
1712 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1715 * Scale scan rate increases based on sharing. There is an
1716 * inverse relationship between the degree of sharing and
1717 * the adjustment made to the scanning period. Broadly
1718 * speaking the intent is that there is little point
1719 * scanning faster if shared accesses dominate as it may
1720 * simply bounce migrations uselessly
1722 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1723 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1726 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1727 task_scan_min(p), task_scan_max(p));
1728 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1732 * Get the fraction of time the task has been running since the last
1733 * NUMA placement cycle. The scheduler keeps similar statistics, but
1734 * decays those on a 32ms period, which is orders of magnitude off
1735 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1736 * stats only if the task is so new there are no NUMA statistics yet.
1738 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1740 u64 runtime, delta, now;
1741 /* Use the start of this time slice to avoid calculations. */
1742 now = p->se.exec_start;
1743 runtime = p->se.sum_exec_runtime;
1745 if (p->last_task_numa_placement) {
1746 delta = runtime - p->last_sum_exec_runtime;
1747 *period = now - p->last_task_numa_placement;
1749 delta = p->se.avg.load_sum / p->se.load.weight;
1750 *period = LOAD_AVG_MAX;
1753 p->last_sum_exec_runtime = runtime;
1754 p->last_task_numa_placement = now;
1760 * Determine the preferred nid for a task in a numa_group. This needs to
1761 * be done in a way that produces consistent results with group_weight,
1762 * otherwise workloads might not converge.
1764 static int preferred_group_nid(struct task_struct *p, int nid)
1769 /* Direct connections between all NUMA nodes. */
1770 if (sched_numa_topology_type == NUMA_DIRECT)
1774 * On a system with glueless mesh NUMA topology, group_weight
1775 * scores nodes according to the number of NUMA hinting faults on
1776 * both the node itself, and on nearby nodes.
1778 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1779 unsigned long score, max_score = 0;
1780 int node, max_node = nid;
1782 dist = sched_max_numa_distance;
1784 for_each_online_node(node) {
1785 score = group_weight(p, node, dist);
1786 if (score > max_score) {
1795 * Finding the preferred nid in a system with NUMA backplane
1796 * interconnect topology is more involved. The goal is to locate
1797 * tasks from numa_groups near each other in the system, and
1798 * untangle workloads from different sides of the system. This requires
1799 * searching down the hierarchy of node groups, recursively searching
1800 * inside the highest scoring group of nodes. The nodemask tricks
1801 * keep the complexity of the search down.
1803 nodes = node_online_map;
1804 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1805 unsigned long max_faults = 0;
1806 nodemask_t max_group = NODE_MASK_NONE;
1809 /* Are there nodes at this distance from each other? */
1810 if (!find_numa_distance(dist))
1813 for_each_node_mask(a, nodes) {
1814 unsigned long faults = 0;
1815 nodemask_t this_group;
1816 nodes_clear(this_group);
1818 /* Sum group's NUMA faults; includes a==b case. */
1819 for_each_node_mask(b, nodes) {
1820 if (node_distance(a, b) < dist) {
1821 faults += group_faults(p, b);
1822 node_set(b, this_group);
1823 node_clear(b, nodes);
1827 /* Remember the top group. */
1828 if (faults > max_faults) {
1829 max_faults = faults;
1830 max_group = this_group;
1832 * subtle: at the smallest distance there is
1833 * just one node left in each "group", the
1834 * winner is the preferred nid.
1839 /* Next round, evaluate the nodes within max_group. */
1847 static void task_numa_placement(struct task_struct *p)
1849 int seq, nid, max_nid = -1, max_group_nid = -1;
1850 unsigned long max_faults = 0, max_group_faults = 0;
1851 unsigned long fault_types[2] = { 0, 0 };
1852 unsigned long total_faults;
1853 u64 runtime, period;
1854 spinlock_t *group_lock = NULL;
1857 * The p->mm->numa_scan_seq field gets updated without
1858 * exclusive access. Use READ_ONCE() here to ensure
1859 * that the field is read in a single access:
1861 seq = READ_ONCE(p->mm->numa_scan_seq);
1862 if (p->numa_scan_seq == seq)
1864 p->numa_scan_seq = seq;
1865 p->numa_scan_period_max = task_scan_max(p);
1867 total_faults = p->numa_faults_locality[0] +
1868 p->numa_faults_locality[1];
1869 runtime = numa_get_avg_runtime(p, &period);
1871 /* If the task is part of a group prevent parallel updates to group stats */
1872 if (p->numa_group) {
1873 group_lock = &p->numa_group->lock;
1874 spin_lock_irq(group_lock);
1877 /* Find the node with the highest number of faults */
1878 for_each_online_node(nid) {
1879 /* Keep track of the offsets in numa_faults array */
1880 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1881 unsigned long faults = 0, group_faults = 0;
1884 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1885 long diff, f_diff, f_weight;
1887 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1888 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1889 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1890 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1892 /* Decay existing window, copy faults since last scan */
1893 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1894 fault_types[priv] += p->numa_faults[membuf_idx];
1895 p->numa_faults[membuf_idx] = 0;
1898 * Normalize the faults_from, so all tasks in a group
1899 * count according to CPU use, instead of by the raw
1900 * number of faults. Tasks with little runtime have
1901 * little over-all impact on throughput, and thus their
1902 * faults are less important.
1904 f_weight = div64_u64(runtime << 16, period + 1);
1905 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1907 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1908 p->numa_faults[cpubuf_idx] = 0;
1910 p->numa_faults[mem_idx] += diff;
1911 p->numa_faults[cpu_idx] += f_diff;
1912 faults += p->numa_faults[mem_idx];
1913 p->total_numa_faults += diff;
1914 if (p->numa_group) {
1916 * safe because we can only change our own group
1918 * mem_idx represents the offset for a given
1919 * nid and priv in a specific region because it
1920 * is at the beginning of the numa_faults array.
1922 p->numa_group->faults[mem_idx] += diff;
1923 p->numa_group->faults_cpu[mem_idx] += f_diff;
1924 p->numa_group->total_faults += diff;
1925 group_faults += p->numa_group->faults[mem_idx];
1929 if (faults > max_faults) {
1930 max_faults = faults;
1934 if (group_faults > max_group_faults) {
1935 max_group_faults = group_faults;
1936 max_group_nid = nid;
1940 update_task_scan_period(p, fault_types[0], fault_types[1]);
1942 if (p->numa_group) {
1943 update_numa_active_node_mask(p->numa_group);
1944 spin_unlock_irq(group_lock);
1945 max_nid = preferred_group_nid(p, max_group_nid);
1949 /* Set the new preferred node */
1950 if (max_nid != p->numa_preferred_nid)
1951 sched_setnuma(p, max_nid);
1953 if (task_node(p) != p->numa_preferred_nid)
1954 numa_migrate_preferred(p);
1958 static inline int get_numa_group(struct numa_group *grp)
1960 return atomic_inc_not_zero(&grp->refcount);
1963 static inline void put_numa_group(struct numa_group *grp)
1965 if (atomic_dec_and_test(&grp->refcount))
1966 kfree_rcu(grp, rcu);
1969 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1972 struct numa_group *grp, *my_grp;
1973 struct task_struct *tsk;
1975 int cpu = cpupid_to_cpu(cpupid);
1978 if (unlikely(!p->numa_group)) {
1979 unsigned int size = sizeof(struct numa_group) +
1980 4*nr_node_ids*sizeof(unsigned long);
1982 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1986 atomic_set(&grp->refcount, 1);
1987 spin_lock_init(&grp->lock);
1989 /* Second half of the array tracks nids where faults happen */
1990 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1993 node_set(task_node(current), grp->active_nodes);
1995 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1996 grp->faults[i] = p->numa_faults[i];
1998 grp->total_faults = p->total_numa_faults;
2001 rcu_assign_pointer(p->numa_group, grp);
2005 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2007 if (!cpupid_match_pid(tsk, cpupid))
2010 grp = rcu_dereference(tsk->numa_group);
2014 my_grp = p->numa_group;
2019 * Only join the other group if its bigger; if we're the bigger group,
2020 * the other task will join us.
2022 if (my_grp->nr_tasks > grp->nr_tasks)
2026 * Tie-break on the grp address.
2028 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2031 /* Always join threads in the same process. */
2032 if (tsk->mm == current->mm)
2035 /* Simple filter to avoid false positives due to PID collisions */
2036 if (flags & TNF_SHARED)
2039 /* Update priv based on whether false sharing was detected */
2042 if (join && !get_numa_group(grp))
2050 BUG_ON(irqs_disabled());
2051 double_lock_irq(&my_grp->lock, &grp->lock);
2053 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2054 my_grp->faults[i] -= p->numa_faults[i];
2055 grp->faults[i] += p->numa_faults[i];
2057 my_grp->total_faults -= p->total_numa_faults;
2058 grp->total_faults += p->total_numa_faults;
2063 spin_unlock(&my_grp->lock);
2064 spin_unlock_irq(&grp->lock);
2066 rcu_assign_pointer(p->numa_group, grp);
2068 put_numa_group(my_grp);
2076 void task_numa_free(struct task_struct *p)
2078 struct numa_group *grp = p->numa_group;
2079 void *numa_faults = p->numa_faults;
2080 unsigned long flags;
2084 spin_lock_irqsave(&grp->lock, flags);
2085 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2086 grp->faults[i] -= p->numa_faults[i];
2087 grp->total_faults -= p->total_numa_faults;
2090 spin_unlock_irqrestore(&grp->lock, flags);
2091 RCU_INIT_POINTER(p->numa_group, NULL);
2092 put_numa_group(grp);
2095 p->numa_faults = NULL;
2100 * Got a PROT_NONE fault for a page on @node.
2102 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2104 struct task_struct *p = current;
2105 bool migrated = flags & TNF_MIGRATED;
2106 int cpu_node = task_node(current);
2107 int local = !!(flags & TNF_FAULT_LOCAL);
2110 if (!static_branch_likely(&sched_numa_balancing))
2113 /* for example, ksmd faulting in a user's mm */
2117 /* Allocate buffer to track faults on a per-node basis */
2118 if (unlikely(!p->numa_faults)) {
2119 int size = sizeof(*p->numa_faults) *
2120 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2122 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2123 if (!p->numa_faults)
2126 p->total_numa_faults = 0;
2127 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2131 * First accesses are treated as private, otherwise consider accesses
2132 * to be private if the accessing pid has not changed
2134 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2137 priv = cpupid_match_pid(p, last_cpupid);
2138 if (!priv && !(flags & TNF_NO_GROUP))
2139 task_numa_group(p, last_cpupid, flags, &priv);
2143 * If a workload spans multiple NUMA nodes, a shared fault that
2144 * occurs wholly within the set of nodes that the workload is
2145 * actively using should be counted as local. This allows the
2146 * scan rate to slow down when a workload has settled down.
2148 if (!priv && !local && p->numa_group &&
2149 node_isset(cpu_node, p->numa_group->active_nodes) &&
2150 node_isset(mem_node, p->numa_group->active_nodes))
2153 task_numa_placement(p);
2156 * Retry task to preferred node migration periodically, in case it
2157 * case it previously failed, or the scheduler moved us.
2159 if (time_after(jiffies, p->numa_migrate_retry))
2160 numa_migrate_preferred(p);
2163 p->numa_pages_migrated += pages;
2164 if (flags & TNF_MIGRATE_FAIL)
2165 p->numa_faults_locality[2] += pages;
2167 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2168 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2169 p->numa_faults_locality[local] += pages;
2172 static void reset_ptenuma_scan(struct task_struct *p)
2175 * We only did a read acquisition of the mmap sem, so
2176 * p->mm->numa_scan_seq is written to without exclusive access
2177 * and the update is not guaranteed to be atomic. That's not
2178 * much of an issue though, since this is just used for
2179 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2180 * expensive, to avoid any form of compiler optimizations:
2182 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2183 p->mm->numa_scan_offset = 0;
2187 * The expensive part of numa migration is done from task_work context.
2188 * Triggered from task_tick_numa().
2190 void task_numa_work(struct callback_head *work)
2192 unsigned long migrate, next_scan, now = jiffies;
2193 struct task_struct *p = current;
2194 struct mm_struct *mm = p->mm;
2195 struct vm_area_struct *vma;
2196 unsigned long start, end;
2197 unsigned long nr_pte_updates = 0;
2198 long pages, virtpages;
2200 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2202 work->next = work; /* protect against double add */
2204 * Who cares about NUMA placement when they're dying.
2206 * NOTE: make sure not to dereference p->mm before this check,
2207 * exit_task_work() happens _after_ exit_mm() so we could be called
2208 * without p->mm even though we still had it when we enqueued this
2211 if (p->flags & PF_EXITING)
2214 if (!mm->numa_next_scan) {
2215 mm->numa_next_scan = now +
2216 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2220 * Enforce maximal scan/migration frequency..
2222 migrate = mm->numa_next_scan;
2223 if (time_before(now, migrate))
2226 if (p->numa_scan_period == 0) {
2227 p->numa_scan_period_max = task_scan_max(p);
2228 p->numa_scan_period = task_scan_min(p);
2231 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2232 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2236 * Delay this task enough that another task of this mm will likely win
2237 * the next time around.
2239 p->node_stamp += 2 * TICK_NSEC;
2241 start = mm->numa_scan_offset;
2242 pages = sysctl_numa_balancing_scan_size;
2243 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2244 virtpages = pages * 8; /* Scan up to this much virtual space */
2249 down_read(&mm->mmap_sem);
2250 vma = find_vma(mm, start);
2252 reset_ptenuma_scan(p);
2256 for (; vma; vma = vma->vm_next) {
2257 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2258 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2263 * Shared library pages mapped by multiple processes are not
2264 * migrated as it is expected they are cache replicated. Avoid
2265 * hinting faults in read-only file-backed mappings or the vdso
2266 * as migrating the pages will be of marginal benefit.
2269 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2273 * Skip inaccessible VMAs to avoid any confusion between
2274 * PROT_NONE and NUMA hinting ptes
2276 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2280 start = max(start, vma->vm_start);
2281 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2282 end = min(end, vma->vm_end);
2283 nr_pte_updates = change_prot_numa(vma, start, end);
2286 * Try to scan sysctl_numa_balancing_size worth of
2287 * hpages that have at least one present PTE that
2288 * is not already pte-numa. If the VMA contains
2289 * areas that are unused or already full of prot_numa
2290 * PTEs, scan up to virtpages, to skip through those
2294 pages -= (end - start) >> PAGE_SHIFT;
2295 virtpages -= (end - start) >> PAGE_SHIFT;
2298 if (pages <= 0 || virtpages <= 0)
2302 } while (end != vma->vm_end);
2307 * It is possible to reach the end of the VMA list but the last few
2308 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2309 * would find the !migratable VMA on the next scan but not reset the
2310 * scanner to the start so check it now.
2313 mm->numa_scan_offset = start;
2315 reset_ptenuma_scan(p);
2316 up_read(&mm->mmap_sem);
2320 * Drive the periodic memory faults..
2322 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2324 struct callback_head *work = &curr->numa_work;
2328 * We don't care about NUMA placement if we don't have memory.
2330 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2334 * Using runtime rather than walltime has the dual advantage that
2335 * we (mostly) drive the selection from busy threads and that the
2336 * task needs to have done some actual work before we bother with
2339 now = curr->se.sum_exec_runtime;
2340 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2342 if (now > curr->node_stamp + period) {
2343 if (!curr->node_stamp)
2344 curr->numa_scan_period = task_scan_min(curr);
2345 curr->node_stamp += period;
2347 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2348 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2349 task_work_add(curr, work, true);
2354 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2358 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2362 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2365 #endif /* CONFIG_NUMA_BALANCING */
2368 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2370 update_load_add(&cfs_rq->load, se->load.weight);
2371 if (!parent_entity(se))
2372 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2374 if (entity_is_task(se)) {
2375 struct rq *rq = rq_of(cfs_rq);
2377 account_numa_enqueue(rq, task_of(se));
2378 list_add(&se->group_node, &rq->cfs_tasks);
2381 cfs_rq->nr_running++;
2385 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2387 update_load_sub(&cfs_rq->load, se->load.weight);
2388 if (!parent_entity(se))
2389 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2390 if (entity_is_task(se)) {
2391 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2392 list_del_init(&se->group_node);
2394 cfs_rq->nr_running--;
2397 #ifdef CONFIG_FAIR_GROUP_SCHED
2399 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2404 * Use this CPU's real-time load instead of the last load contribution
2405 * as the updating of the contribution is delayed, and we will use the
2406 * the real-time load to calc the share. See update_tg_load_avg().
2408 tg_weight = atomic_long_read(&tg->load_avg);
2409 tg_weight -= cfs_rq->tg_load_avg_contrib;
2410 tg_weight += cfs_rq->load.weight;
2415 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2417 long tg_weight, load, shares;
2419 tg_weight = calc_tg_weight(tg, cfs_rq);
2420 load = cfs_rq->load.weight;
2422 shares = (tg->shares * load);
2424 shares /= tg_weight;
2426 if (shares < MIN_SHARES)
2427 shares = MIN_SHARES;
2428 if (shares > tg->shares)
2429 shares = tg->shares;
2433 # else /* CONFIG_SMP */
2434 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2438 # endif /* CONFIG_SMP */
2439 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2440 unsigned long weight)
2443 /* commit outstanding execution time */
2444 if (cfs_rq->curr == se)
2445 update_curr(cfs_rq);
2446 account_entity_dequeue(cfs_rq, se);
2449 update_load_set(&se->load, weight);
2452 account_entity_enqueue(cfs_rq, se);
2455 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2457 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2459 struct task_group *tg;
2460 struct sched_entity *se;
2464 se = tg->se[cpu_of(rq_of(cfs_rq))];
2465 if (!se || throttled_hierarchy(cfs_rq))
2468 if (likely(se->load.weight == tg->shares))
2471 shares = calc_cfs_shares(cfs_rq, tg);
2473 reweight_entity(cfs_rq_of(se), se, shares);
2475 #else /* CONFIG_FAIR_GROUP_SCHED */
2476 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2479 #endif /* CONFIG_FAIR_GROUP_SCHED */
2482 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2483 static const u32 runnable_avg_yN_inv[] = {
2484 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2485 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2486 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2487 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2488 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2489 0x85aac367, 0x82cd8698,
2493 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2494 * over-estimates when re-combining.
2496 static const u32 runnable_avg_yN_sum[] = {
2497 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2498 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2499 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2504 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2506 static __always_inline u64 decay_load(u64 val, u64 n)
2508 unsigned int local_n;
2512 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2515 /* after bounds checking we can collapse to 32-bit */
2519 * As y^PERIOD = 1/2, we can combine
2520 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2521 * With a look-up table which covers y^n (n<PERIOD)
2523 * To achieve constant time decay_load.
2525 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2526 val >>= local_n / LOAD_AVG_PERIOD;
2527 local_n %= LOAD_AVG_PERIOD;
2530 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2535 * For updates fully spanning n periods, the contribution to runnable
2536 * average will be: \Sum 1024*y^n
2538 * We can compute this reasonably efficiently by combining:
2539 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2541 static u32 __compute_runnable_contrib(u64 n)
2545 if (likely(n <= LOAD_AVG_PERIOD))
2546 return runnable_avg_yN_sum[n];
2547 else if (unlikely(n >= LOAD_AVG_MAX_N))
2548 return LOAD_AVG_MAX;
2550 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2552 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2553 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2555 n -= LOAD_AVG_PERIOD;
2556 } while (n > LOAD_AVG_PERIOD);
2558 contrib = decay_load(contrib, n);
2559 return contrib + runnable_avg_yN_sum[n];
2562 #if (SCHED_LOAD_SHIFT - SCHED_LOAD_RESOLUTION) != 10 || SCHED_CAPACITY_SHIFT != 10
2563 #error "load tracking assumes 2^10 as unit"
2566 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2569 * We can represent the historical contribution to runnable average as the
2570 * coefficients of a geometric series. To do this we sub-divide our runnable
2571 * history into segments of approximately 1ms (1024us); label the segment that
2572 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2574 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2576 * (now) (~1ms ago) (~2ms ago)
2578 * Let u_i denote the fraction of p_i that the entity was runnable.
2580 * We then designate the fractions u_i as our co-efficients, yielding the
2581 * following representation of historical load:
2582 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2584 * We choose y based on the with of a reasonably scheduling period, fixing:
2587 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2588 * approximately half as much as the contribution to load within the last ms
2591 * When a period "rolls over" and we have new u_0`, multiplying the previous
2592 * sum again by y is sufficient to update:
2593 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2594 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2596 static __always_inline int
2597 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2598 unsigned long weight, int running, struct cfs_rq *cfs_rq)
2600 u64 delta, scaled_delta, periods;
2602 unsigned int delta_w, scaled_delta_w, decayed = 0;
2603 unsigned long scale_freq, scale_cpu;
2605 delta = now - sa->last_update_time;
2607 * This should only happen when time goes backwards, which it
2608 * unfortunately does during sched clock init when we swap over to TSC.
2610 if ((s64)delta < 0) {
2611 sa->last_update_time = now;
2616 * Use 1024ns as the unit of measurement since it's a reasonable
2617 * approximation of 1us and fast to compute.
2622 sa->last_update_time = now;
2624 scale_freq = arch_scale_freq_capacity(NULL, cpu);
2625 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2626 trace_sched_contrib_scale_f(cpu, scale_freq, scale_cpu);
2628 /* delta_w is the amount already accumulated against our next period */
2629 delta_w = sa->period_contrib;
2630 if (delta + delta_w >= 1024) {
2633 /* how much left for next period will start over, we don't know yet */
2634 sa->period_contrib = 0;
2637 * Now that we know we're crossing a period boundary, figure
2638 * out how much from delta we need to complete the current
2639 * period and accrue it.
2641 delta_w = 1024 - delta_w;
2642 scaled_delta_w = cap_scale(delta_w, scale_freq);
2644 sa->load_sum += weight * scaled_delta_w;
2646 cfs_rq->runnable_load_sum +=
2647 weight * scaled_delta_w;
2651 sa->util_sum += scaled_delta_w * scale_cpu;
2655 /* Figure out how many additional periods this update spans */
2656 periods = delta / 1024;
2659 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2661 cfs_rq->runnable_load_sum =
2662 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2664 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2666 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2667 contrib = __compute_runnable_contrib(periods);
2668 contrib = cap_scale(contrib, scale_freq);
2670 sa->load_sum += weight * contrib;
2672 cfs_rq->runnable_load_sum += weight * contrib;
2675 sa->util_sum += contrib * scale_cpu;
2678 /* Remainder of delta accrued against u_0` */
2679 scaled_delta = cap_scale(delta, scale_freq);
2681 sa->load_sum += weight * scaled_delta;
2683 cfs_rq->runnable_load_sum += weight * scaled_delta;
2686 sa->util_sum += scaled_delta * scale_cpu;
2688 sa->period_contrib += delta;
2691 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2693 cfs_rq->runnable_load_avg =
2694 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2696 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2702 #ifdef CONFIG_FAIR_GROUP_SCHED
2704 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2705 * and effective_load (which is not done because it is too costly).
2707 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2709 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2711 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2712 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2713 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2717 #else /* CONFIG_FAIR_GROUP_SCHED */
2718 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2719 #endif /* CONFIG_FAIR_GROUP_SCHED */
2721 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2723 if (&this_rq()->cfs == cfs_rq) {
2725 * There are a few boundary cases this might miss but it should
2726 * get called often enough that that should (hopefully) not be
2727 * a real problem -- added to that it only calls on the local
2728 * CPU, so if we enqueue remotely we'll miss an update, but
2729 * the next tick/schedule should update.
2731 * It will not get called when we go idle, because the idle
2732 * thread is a different class (!fair), nor will the utilization
2733 * number include things like RT tasks.
2735 * As is, the util number is not freq-invariant (we'd have to
2736 * implement arch_scale_freq_capacity() for that).
2740 cpufreq_update_util(rq_of(cfs_rq), 0);
2744 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2747 * Unsigned subtract and clamp on underflow.
2749 * Explicitly do a load-store to ensure the intermediate value never hits
2750 * memory. This allows lockless observations without ever seeing the negative
2753 #define sub_positive(_ptr, _val) do { \
2754 typeof(_ptr) ptr = (_ptr); \
2755 typeof(*ptr) val = (_val); \
2756 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2760 WRITE_ONCE(*ptr, res); \
2763 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2764 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq,
2767 struct sched_avg *sa = &cfs_rq->avg;
2768 int decayed, removed = 0, removed_util = 0;
2770 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2771 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2772 sub_positive(&sa->load_avg, r);
2773 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2777 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2778 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2779 sub_positive(&sa->util_avg, r);
2780 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2784 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2785 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2787 #ifndef CONFIG_64BIT
2789 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2792 /* Trace CPU load, unless cfs_rq belongs to a non-root task_group */
2793 if (cfs_rq == &rq_of(cfs_rq)->cfs)
2794 trace_sched_load_avg_cpu(cpu_of(rq_of(cfs_rq)), cfs_rq);
2796 if (update_freq && (decayed || removed_util))
2797 cfs_rq_util_change(cfs_rq);
2799 return decayed || removed;
2802 /* Update task and its cfs_rq load average */
2803 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2805 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2806 u64 now = cfs_rq_clock_task(cfs_rq);
2807 int cpu = cpu_of(rq_of(cfs_rq));
2810 * Track task load average for carrying it to new CPU after migrated, and
2811 * track group sched_entity load average for task_h_load calc in migration
2813 __update_load_avg(now, cpu, &se->avg,
2814 se->on_rq * scale_load_down(se->load.weight),
2815 cfs_rq->curr == se, NULL);
2817 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2818 update_tg_load_avg(cfs_rq, 0);
2820 if (entity_is_task(se))
2821 trace_sched_load_avg_task(task_of(se), &se->avg);
2824 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2826 if (!sched_feat(ATTACH_AGE_LOAD))
2830 * If we got migrated (either between CPUs or between cgroups) we'll
2831 * have aged the average right before clearing @last_update_time.
2833 if (se->avg.last_update_time) {
2834 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2835 &se->avg, 0, 0, NULL);
2838 * XXX: we could have just aged the entire load away if we've been
2839 * absent from the fair class for too long.
2844 se->avg.last_update_time = cfs_rq->avg.last_update_time;
2845 cfs_rq->avg.load_avg += se->avg.load_avg;
2846 cfs_rq->avg.load_sum += se->avg.load_sum;
2847 cfs_rq->avg.util_avg += se->avg.util_avg;
2848 cfs_rq->avg.util_sum += se->avg.util_sum;
2850 cfs_rq_util_change(cfs_rq);
2853 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2855 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
2856 &se->avg, se->on_rq * scale_load_down(se->load.weight),
2857 cfs_rq->curr == se, NULL);
2859 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2860 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
2861 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
2862 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
2864 cfs_rq_util_change(cfs_rq);
2867 /* Add the load generated by se into cfs_rq's load average */
2869 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2871 struct sched_avg *sa = &se->avg;
2872 u64 now = cfs_rq_clock_task(cfs_rq);
2873 int migrated, decayed;
2875 migrated = !sa->last_update_time;
2877 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2878 se->on_rq * scale_load_down(se->load.weight),
2879 cfs_rq->curr == se, NULL);
2882 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
2884 cfs_rq->runnable_load_avg += sa->load_avg;
2885 cfs_rq->runnable_load_sum += sa->load_sum;
2888 attach_entity_load_avg(cfs_rq, se);
2890 if (decayed || migrated)
2891 update_tg_load_avg(cfs_rq, 0);
2894 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
2896 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2898 update_load_avg(se, 1);
2900 cfs_rq->runnable_load_avg =
2901 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
2902 cfs_rq->runnable_load_sum =
2903 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
2906 #ifndef CONFIG_64BIT
2907 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2909 u64 last_update_time_copy;
2910 u64 last_update_time;
2913 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2915 last_update_time = cfs_rq->avg.last_update_time;
2916 } while (last_update_time != last_update_time_copy);
2918 return last_update_time;
2921 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
2923 return cfs_rq->avg.last_update_time;
2928 * Synchronize entity load avg of dequeued entity without locking
2931 void sync_entity_load_avg(struct sched_entity *se)
2933 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2934 u64 last_update_time;
2936 last_update_time = cfs_rq_last_update_time(cfs_rq);
2937 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
2941 * Task first catches up with cfs_rq, and then subtract
2942 * itself from the cfs_rq (task must be off the queue now).
2944 void remove_entity_load_avg(struct sched_entity *se)
2946 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2949 * Newly created task or never used group entity should not be removed
2950 * from its (source) cfs_rq
2952 if (se->avg.last_update_time == 0)
2955 sync_entity_load_avg(se);
2956 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2957 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2961 * Update the rq's load with the elapsed running time before entering
2962 * idle. if the last scheduled task is not a CFS task, idle_enter will
2963 * be the only way to update the runnable statistic.
2965 void idle_enter_fair(struct rq *this_rq)
2970 * Update the rq's load with the elapsed idle time before a task is
2971 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2972 * be the only way to update the runnable statistic.
2974 void idle_exit_fair(struct rq *this_rq)
2978 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
2980 return cfs_rq->runnable_load_avg;
2983 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
2985 return cfs_rq->avg.load_avg;
2988 static int idle_balance(struct rq *this_rq);
2990 #else /* CONFIG_SMP */
2992 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2994 cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
2998 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3000 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3001 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3004 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3006 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3008 static inline int idle_balance(struct rq *rq)
3013 #endif /* CONFIG_SMP */
3015 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3017 #ifdef CONFIG_SCHEDSTATS
3018 struct task_struct *tsk = NULL;
3020 if (entity_is_task(se))
3023 if (se->statistics.sleep_start) {
3024 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3029 if (unlikely(delta > se->statistics.sleep_max))
3030 se->statistics.sleep_max = delta;
3032 se->statistics.sleep_start = 0;
3033 se->statistics.sum_sleep_runtime += delta;
3036 account_scheduler_latency(tsk, delta >> 10, 1);
3037 trace_sched_stat_sleep(tsk, delta);
3040 if (se->statistics.block_start) {
3041 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3046 if (unlikely(delta > se->statistics.block_max))
3047 se->statistics.block_max = delta;
3049 se->statistics.block_start = 0;
3050 se->statistics.sum_sleep_runtime += delta;
3053 if (tsk->in_iowait) {
3054 se->statistics.iowait_sum += delta;
3055 se->statistics.iowait_count++;
3056 trace_sched_stat_iowait(tsk, delta);
3059 trace_sched_stat_blocked(tsk, delta);
3060 trace_sched_blocked_reason(tsk);
3063 * Blocking time is in units of nanosecs, so shift by
3064 * 20 to get a milliseconds-range estimation of the
3065 * amount of time that the task spent sleeping:
3067 if (unlikely(prof_on == SLEEP_PROFILING)) {
3068 profile_hits(SLEEP_PROFILING,
3069 (void *)get_wchan(tsk),
3072 account_scheduler_latency(tsk, delta >> 10, 0);
3078 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3080 #ifdef CONFIG_SCHED_DEBUG
3081 s64 d = se->vruntime - cfs_rq->min_vruntime;
3086 if (d > 3*sysctl_sched_latency)
3087 schedstat_inc(cfs_rq, nr_spread_over);
3092 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3094 u64 vruntime = cfs_rq->min_vruntime;
3097 * The 'current' period is already promised to the current tasks,
3098 * however the extra weight of the new task will slow them down a
3099 * little, place the new task so that it fits in the slot that
3100 * stays open at the end.
3102 if (initial && sched_feat(START_DEBIT))
3103 vruntime += sched_vslice(cfs_rq, se);
3105 /* sleeps up to a single latency don't count. */
3107 unsigned long thresh = sysctl_sched_latency;
3110 * Halve their sleep time's effect, to allow
3111 * for a gentler effect of sleepers:
3113 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3119 /* ensure we never gain time by being placed backwards. */
3120 se->vruntime = max_vruntime(se->vruntime, vruntime);
3123 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3126 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3129 * Update the normalized vruntime before updating min_vruntime
3130 * through calling update_curr().
3132 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
3133 se->vruntime += cfs_rq->min_vruntime;
3136 * Update run-time statistics of the 'current'.
3138 update_curr(cfs_rq);
3139 enqueue_entity_load_avg(cfs_rq, se);
3140 account_entity_enqueue(cfs_rq, se);
3141 update_cfs_shares(cfs_rq);
3143 if (flags & ENQUEUE_WAKEUP) {
3144 place_entity(cfs_rq, se, 0);
3145 enqueue_sleeper(cfs_rq, se);
3148 update_stats_enqueue(cfs_rq, se);
3149 check_spread(cfs_rq, se);
3150 if (se != cfs_rq->curr)
3151 __enqueue_entity(cfs_rq, se);
3154 if (cfs_rq->nr_running == 1) {
3155 list_add_leaf_cfs_rq(cfs_rq);
3156 check_enqueue_throttle(cfs_rq);
3160 static void __clear_buddies_last(struct sched_entity *se)
3162 for_each_sched_entity(se) {
3163 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3164 if (cfs_rq->last != se)
3167 cfs_rq->last = NULL;
3171 static void __clear_buddies_next(struct sched_entity *se)
3173 for_each_sched_entity(se) {
3174 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3175 if (cfs_rq->next != se)
3178 cfs_rq->next = NULL;
3182 static void __clear_buddies_skip(struct sched_entity *se)
3184 for_each_sched_entity(se) {
3185 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3186 if (cfs_rq->skip != se)
3189 cfs_rq->skip = NULL;
3193 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3195 if (cfs_rq->last == se)
3196 __clear_buddies_last(se);
3198 if (cfs_rq->next == se)
3199 __clear_buddies_next(se);
3201 if (cfs_rq->skip == se)
3202 __clear_buddies_skip(se);
3205 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3208 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3211 * Update run-time statistics of the 'current'.
3213 update_curr(cfs_rq);
3214 dequeue_entity_load_avg(cfs_rq, se);
3216 update_stats_dequeue(cfs_rq, se);
3217 if (flags & DEQUEUE_SLEEP) {
3218 #ifdef CONFIG_SCHEDSTATS
3219 if (entity_is_task(se)) {
3220 struct task_struct *tsk = task_of(se);
3222 if (tsk->state & TASK_INTERRUPTIBLE)
3223 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
3224 if (tsk->state & TASK_UNINTERRUPTIBLE)
3225 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
3230 clear_buddies(cfs_rq, se);
3232 if (se != cfs_rq->curr)
3233 __dequeue_entity(cfs_rq, se);
3235 account_entity_dequeue(cfs_rq, se);
3238 * Normalize the entity after updating the min_vruntime because the
3239 * update can refer to the ->curr item and we need to reflect this
3240 * movement in our normalized position.
3242 if (!(flags & DEQUEUE_SLEEP))
3243 se->vruntime -= cfs_rq->min_vruntime;
3245 /* return excess runtime on last dequeue */
3246 return_cfs_rq_runtime(cfs_rq);
3248 update_min_vruntime(cfs_rq);
3249 update_cfs_shares(cfs_rq);
3253 * Preempt the current task with a newly woken task if needed:
3256 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3258 unsigned long ideal_runtime, delta_exec;
3259 struct sched_entity *se;
3262 ideal_runtime = sched_slice(cfs_rq, curr);
3263 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3264 if (delta_exec > ideal_runtime) {
3265 resched_curr(rq_of(cfs_rq));
3267 * The current task ran long enough, ensure it doesn't get
3268 * re-elected due to buddy favours.
3270 clear_buddies(cfs_rq, curr);
3275 * Ensure that a task that missed wakeup preemption by a
3276 * narrow margin doesn't have to wait for a full slice.
3277 * This also mitigates buddy induced latencies under load.
3279 if (delta_exec < sysctl_sched_min_granularity)
3282 se = __pick_first_entity(cfs_rq);
3283 delta = curr->vruntime - se->vruntime;
3288 if (delta > ideal_runtime)
3289 resched_curr(rq_of(cfs_rq));
3293 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3295 /* 'current' is not kept within the tree. */
3298 * Any task has to be enqueued before it get to execute on
3299 * a CPU. So account for the time it spent waiting on the
3302 update_stats_wait_end(cfs_rq, se);
3303 __dequeue_entity(cfs_rq, se);
3304 update_load_avg(se, 1);
3307 update_stats_curr_start(cfs_rq, se);
3309 #ifdef CONFIG_SCHEDSTATS
3311 * Track our maximum slice length, if the CPU's load is at
3312 * least twice that of our own weight (i.e. dont track it
3313 * when there are only lesser-weight tasks around):
3315 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3316 se->statistics.slice_max = max(se->statistics.slice_max,
3317 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3320 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3324 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3327 * Pick the next process, keeping these things in mind, in this order:
3328 * 1) keep things fair between processes/task groups
3329 * 2) pick the "next" process, since someone really wants that to run
3330 * 3) pick the "last" process, for cache locality
3331 * 4) do not run the "skip" process, if something else is available
3333 static struct sched_entity *
3334 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3336 struct sched_entity *left = __pick_first_entity(cfs_rq);
3337 struct sched_entity *se;
3340 * If curr is set we have to see if its left of the leftmost entity
3341 * still in the tree, provided there was anything in the tree at all.
3343 if (!left || (curr && entity_before(curr, left)))
3346 se = left; /* ideally we run the leftmost entity */
3349 * Avoid running the skip buddy, if running something else can
3350 * be done without getting too unfair.
3352 if (cfs_rq->skip == se) {
3353 struct sched_entity *second;
3356 second = __pick_first_entity(cfs_rq);
3358 second = __pick_next_entity(se);
3359 if (!second || (curr && entity_before(curr, second)))
3363 if (second && wakeup_preempt_entity(second, left) < 1)
3368 * Prefer last buddy, try to return the CPU to a preempted task.
3370 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3374 * Someone really wants this to run. If it's not unfair, run it.
3376 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3379 clear_buddies(cfs_rq, se);
3384 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3386 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3389 * If still on the runqueue then deactivate_task()
3390 * was not called and update_curr() has to be done:
3393 update_curr(cfs_rq);
3395 /* throttle cfs_rqs exceeding runtime */
3396 check_cfs_rq_runtime(cfs_rq);
3398 check_spread(cfs_rq, prev);
3400 update_stats_wait_start(cfs_rq, prev);
3401 /* Put 'current' back into the tree. */
3402 __enqueue_entity(cfs_rq, prev);
3403 /* in !on_rq case, update occurred at dequeue */
3404 update_load_avg(prev, 0);
3406 cfs_rq->curr = NULL;
3410 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3413 * Update run-time statistics of the 'current'.
3415 update_curr(cfs_rq);
3418 * Ensure that runnable average is periodically updated.
3420 update_load_avg(curr, 1);
3421 update_cfs_shares(cfs_rq);
3423 #ifdef CONFIG_SCHED_HRTICK
3425 * queued ticks are scheduled to match the slice, so don't bother
3426 * validating it and just reschedule.
3429 resched_curr(rq_of(cfs_rq));
3433 * don't let the period tick interfere with the hrtick preemption
3435 if (!sched_feat(DOUBLE_TICK) &&
3436 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3440 if (cfs_rq->nr_running > 1)
3441 check_preempt_tick(cfs_rq, curr);
3445 /**************************************************
3446 * CFS bandwidth control machinery
3449 #ifdef CONFIG_CFS_BANDWIDTH
3451 #ifdef HAVE_JUMP_LABEL
3452 static struct static_key __cfs_bandwidth_used;
3454 static inline bool cfs_bandwidth_used(void)
3456 return static_key_false(&__cfs_bandwidth_used);
3459 void cfs_bandwidth_usage_inc(void)
3461 static_key_slow_inc(&__cfs_bandwidth_used);
3464 void cfs_bandwidth_usage_dec(void)
3466 static_key_slow_dec(&__cfs_bandwidth_used);
3468 #else /* HAVE_JUMP_LABEL */
3469 static bool cfs_bandwidth_used(void)
3474 void cfs_bandwidth_usage_inc(void) {}
3475 void cfs_bandwidth_usage_dec(void) {}
3476 #endif /* HAVE_JUMP_LABEL */
3479 * default period for cfs group bandwidth.
3480 * default: 0.1s, units: nanoseconds
3482 static inline u64 default_cfs_period(void)
3484 return 100000000ULL;
3487 static inline u64 sched_cfs_bandwidth_slice(void)
3489 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3493 * Replenish runtime according to assigned quota and update expiration time.
3494 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3495 * additional synchronization around rq->lock.
3497 * requires cfs_b->lock
3499 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3503 if (cfs_b->quota == RUNTIME_INF)
3506 now = sched_clock_cpu(smp_processor_id());
3507 cfs_b->runtime = cfs_b->quota;
3508 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3511 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3513 return &tg->cfs_bandwidth;
3516 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3517 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3519 if (unlikely(cfs_rq->throttle_count))
3520 return cfs_rq->throttled_clock_task;
3522 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3525 /* returns 0 on failure to allocate runtime */
3526 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3528 struct task_group *tg = cfs_rq->tg;
3529 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3530 u64 amount = 0, min_amount, expires;
3532 /* note: this is a positive sum as runtime_remaining <= 0 */
3533 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3535 raw_spin_lock(&cfs_b->lock);
3536 if (cfs_b->quota == RUNTIME_INF)
3537 amount = min_amount;
3539 start_cfs_bandwidth(cfs_b);
3541 if (cfs_b->runtime > 0) {
3542 amount = min(cfs_b->runtime, min_amount);
3543 cfs_b->runtime -= amount;
3547 expires = cfs_b->runtime_expires;
3548 raw_spin_unlock(&cfs_b->lock);
3550 cfs_rq->runtime_remaining += amount;
3552 * we may have advanced our local expiration to account for allowed
3553 * spread between our sched_clock and the one on which runtime was
3556 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3557 cfs_rq->runtime_expires = expires;
3559 return cfs_rq->runtime_remaining > 0;
3563 * Note: This depends on the synchronization provided by sched_clock and the
3564 * fact that rq->clock snapshots this value.
3566 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3568 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3570 /* if the deadline is ahead of our clock, nothing to do */
3571 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3574 if (cfs_rq->runtime_remaining < 0)
3578 * If the local deadline has passed we have to consider the
3579 * possibility that our sched_clock is 'fast' and the global deadline
3580 * has not truly expired.
3582 * Fortunately we can check determine whether this the case by checking
3583 * whether the global deadline has advanced. It is valid to compare
3584 * cfs_b->runtime_expires without any locks since we only care about
3585 * exact equality, so a partial write will still work.
3588 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3589 /* extend local deadline, drift is bounded above by 2 ticks */
3590 cfs_rq->runtime_expires += TICK_NSEC;
3592 /* global deadline is ahead, expiration has passed */
3593 cfs_rq->runtime_remaining = 0;
3597 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3599 /* dock delta_exec before expiring quota (as it could span periods) */
3600 cfs_rq->runtime_remaining -= delta_exec;
3601 expire_cfs_rq_runtime(cfs_rq);
3603 if (likely(cfs_rq->runtime_remaining > 0))
3607 * if we're unable to extend our runtime we resched so that the active
3608 * hierarchy can be throttled
3610 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3611 resched_curr(rq_of(cfs_rq));
3614 static __always_inline
3615 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3617 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3620 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3623 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3625 return cfs_bandwidth_used() && cfs_rq->throttled;
3628 /* check whether cfs_rq, or any parent, is throttled */
3629 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3631 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3635 * Ensure that neither of the group entities corresponding to src_cpu or
3636 * dest_cpu are members of a throttled hierarchy when performing group
3637 * load-balance operations.
3639 static inline int throttled_lb_pair(struct task_group *tg,
3640 int src_cpu, int dest_cpu)
3642 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3644 src_cfs_rq = tg->cfs_rq[src_cpu];
3645 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3647 return throttled_hierarchy(src_cfs_rq) ||
3648 throttled_hierarchy(dest_cfs_rq);
3651 /* updated child weight may affect parent so we have to do this bottom up */
3652 static int tg_unthrottle_up(struct task_group *tg, void *data)
3654 struct rq *rq = data;
3655 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3657 cfs_rq->throttle_count--;
3659 if (!cfs_rq->throttle_count) {
3660 /* adjust cfs_rq_clock_task() */
3661 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3662 cfs_rq->throttled_clock_task;
3669 static int tg_throttle_down(struct task_group *tg, void *data)
3671 struct rq *rq = data;
3672 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3674 /* group is entering throttled state, stop time */
3675 if (!cfs_rq->throttle_count)
3676 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3677 cfs_rq->throttle_count++;
3682 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3684 struct rq *rq = rq_of(cfs_rq);
3685 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3686 struct sched_entity *se;
3687 long task_delta, dequeue = 1;
3690 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3692 /* freeze hierarchy runnable averages while throttled */
3694 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3697 task_delta = cfs_rq->h_nr_running;
3698 for_each_sched_entity(se) {
3699 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3700 /* throttled entity or throttle-on-deactivate */
3705 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3706 qcfs_rq->h_nr_running -= task_delta;
3708 if (qcfs_rq->load.weight)
3713 sub_nr_running(rq, task_delta);
3715 cfs_rq->throttled = 1;
3716 cfs_rq->throttled_clock = rq_clock(rq);
3717 raw_spin_lock(&cfs_b->lock);
3718 empty = list_empty(&cfs_b->throttled_cfs_rq);
3721 * Add to the _head_ of the list, so that an already-started
3722 * distribute_cfs_runtime will not see us
3724 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3727 * If we're the first throttled task, make sure the bandwidth
3731 start_cfs_bandwidth(cfs_b);
3733 raw_spin_unlock(&cfs_b->lock);
3736 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3738 struct rq *rq = rq_of(cfs_rq);
3739 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3740 struct sched_entity *se;
3744 se = cfs_rq->tg->se[cpu_of(rq)];
3746 cfs_rq->throttled = 0;
3748 update_rq_clock(rq);
3750 raw_spin_lock(&cfs_b->lock);
3751 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3752 list_del_rcu(&cfs_rq->throttled_list);
3753 raw_spin_unlock(&cfs_b->lock);
3755 /* update hierarchical throttle state */
3756 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3758 if (!cfs_rq->load.weight)
3761 task_delta = cfs_rq->h_nr_running;
3762 for_each_sched_entity(se) {
3766 cfs_rq = cfs_rq_of(se);
3768 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3769 cfs_rq->h_nr_running += task_delta;
3771 if (cfs_rq_throttled(cfs_rq))
3776 add_nr_running(rq, task_delta);
3778 /* determine whether we need to wake up potentially idle cpu */
3779 if (rq->curr == rq->idle && rq->cfs.nr_running)
3783 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3784 u64 remaining, u64 expires)
3786 struct cfs_rq *cfs_rq;
3788 u64 starting_runtime = remaining;
3791 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3793 struct rq *rq = rq_of(cfs_rq);
3795 raw_spin_lock(&rq->lock);
3796 if (!cfs_rq_throttled(cfs_rq))
3799 runtime = -cfs_rq->runtime_remaining + 1;
3800 if (runtime > remaining)
3801 runtime = remaining;
3802 remaining -= runtime;
3804 cfs_rq->runtime_remaining += runtime;
3805 cfs_rq->runtime_expires = expires;
3807 /* we check whether we're throttled above */
3808 if (cfs_rq->runtime_remaining > 0)
3809 unthrottle_cfs_rq(cfs_rq);
3812 raw_spin_unlock(&rq->lock);
3819 return starting_runtime - remaining;
3823 * Responsible for refilling a task_group's bandwidth and unthrottling its
3824 * cfs_rqs as appropriate. If there has been no activity within the last
3825 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3826 * used to track this state.
3828 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3830 u64 runtime, runtime_expires;
3833 /* no need to continue the timer with no bandwidth constraint */
3834 if (cfs_b->quota == RUNTIME_INF)
3835 goto out_deactivate;
3837 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3838 cfs_b->nr_periods += overrun;
3841 * idle depends on !throttled (for the case of a large deficit), and if
3842 * we're going inactive then everything else can be deferred
3844 if (cfs_b->idle && !throttled)
3845 goto out_deactivate;
3847 __refill_cfs_bandwidth_runtime(cfs_b);
3850 /* mark as potentially idle for the upcoming period */
3855 /* account preceding periods in which throttling occurred */
3856 cfs_b->nr_throttled += overrun;
3858 runtime_expires = cfs_b->runtime_expires;
3861 * This check is repeated as we are holding onto the new bandwidth while
3862 * we unthrottle. This can potentially race with an unthrottled group
3863 * trying to acquire new bandwidth from the global pool. This can result
3864 * in us over-using our runtime if it is all used during this loop, but
3865 * only by limited amounts in that extreme case.
3867 while (throttled && cfs_b->runtime > 0) {
3868 runtime = cfs_b->runtime;
3869 raw_spin_unlock(&cfs_b->lock);
3870 /* we can't nest cfs_b->lock while distributing bandwidth */
3871 runtime = distribute_cfs_runtime(cfs_b, runtime,
3873 raw_spin_lock(&cfs_b->lock);
3875 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3877 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3881 * While we are ensured activity in the period following an
3882 * unthrottle, this also covers the case in which the new bandwidth is
3883 * insufficient to cover the existing bandwidth deficit. (Forcing the
3884 * timer to remain active while there are any throttled entities.)
3894 /* a cfs_rq won't donate quota below this amount */
3895 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3896 /* minimum remaining period time to redistribute slack quota */
3897 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3898 /* how long we wait to gather additional slack before distributing */
3899 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3902 * Are we near the end of the current quota period?
3904 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3905 * hrtimer base being cleared by hrtimer_start. In the case of
3906 * migrate_hrtimers, base is never cleared, so we are fine.
3908 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3910 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3913 /* if the call-back is running a quota refresh is already occurring */
3914 if (hrtimer_callback_running(refresh_timer))
3917 /* is a quota refresh about to occur? */
3918 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3919 if (remaining < min_expire)
3925 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3927 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3929 /* if there's a quota refresh soon don't bother with slack */
3930 if (runtime_refresh_within(cfs_b, min_left))
3933 hrtimer_start(&cfs_b->slack_timer,
3934 ns_to_ktime(cfs_bandwidth_slack_period),
3938 /* we know any runtime found here is valid as update_curr() precedes return */
3939 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3941 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3942 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3944 if (slack_runtime <= 0)
3947 raw_spin_lock(&cfs_b->lock);
3948 if (cfs_b->quota != RUNTIME_INF &&
3949 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3950 cfs_b->runtime += slack_runtime;
3952 /* we are under rq->lock, defer unthrottling using a timer */
3953 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3954 !list_empty(&cfs_b->throttled_cfs_rq))
3955 start_cfs_slack_bandwidth(cfs_b);
3957 raw_spin_unlock(&cfs_b->lock);
3959 /* even if it's not valid for return we don't want to try again */
3960 cfs_rq->runtime_remaining -= slack_runtime;
3963 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3965 if (!cfs_bandwidth_used())
3968 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3971 __return_cfs_rq_runtime(cfs_rq);
3975 * This is done with a timer (instead of inline with bandwidth return) since
3976 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3978 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3980 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3983 /* confirm we're still not at a refresh boundary */
3984 raw_spin_lock(&cfs_b->lock);
3985 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3986 raw_spin_unlock(&cfs_b->lock);
3990 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3991 runtime = cfs_b->runtime;
3993 expires = cfs_b->runtime_expires;
3994 raw_spin_unlock(&cfs_b->lock);
3999 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4001 raw_spin_lock(&cfs_b->lock);
4002 if (expires == cfs_b->runtime_expires)
4003 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4004 raw_spin_unlock(&cfs_b->lock);
4008 * When a group wakes up we want to make sure that its quota is not already
4009 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4010 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4012 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4014 if (!cfs_bandwidth_used())
4017 /* Synchronize hierarchical throttle counter: */
4018 if (unlikely(!cfs_rq->throttle_uptodate)) {
4019 struct rq *rq = rq_of(cfs_rq);
4020 struct cfs_rq *pcfs_rq;
4021 struct task_group *tg;
4023 cfs_rq->throttle_uptodate = 1;
4025 /* Get closest up-to-date node, because leaves go first: */
4026 for (tg = cfs_rq->tg->parent; tg; tg = tg->parent) {
4027 pcfs_rq = tg->cfs_rq[cpu_of(rq)];
4028 if (pcfs_rq->throttle_uptodate)
4032 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4033 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4037 /* an active group must be handled by the update_curr()->put() path */
4038 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4041 /* ensure the group is not already throttled */
4042 if (cfs_rq_throttled(cfs_rq))
4045 /* update runtime allocation */
4046 account_cfs_rq_runtime(cfs_rq, 0);
4047 if (cfs_rq->runtime_remaining <= 0)
4048 throttle_cfs_rq(cfs_rq);
4051 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4052 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4054 if (!cfs_bandwidth_used())
4057 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4061 * it's possible for a throttled entity to be forced into a running
4062 * state (e.g. set_curr_task), in this case we're finished.
4064 if (cfs_rq_throttled(cfs_rq))
4067 throttle_cfs_rq(cfs_rq);
4071 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4073 struct cfs_bandwidth *cfs_b =
4074 container_of(timer, struct cfs_bandwidth, slack_timer);
4076 do_sched_cfs_slack_timer(cfs_b);
4078 return HRTIMER_NORESTART;
4081 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4083 struct cfs_bandwidth *cfs_b =
4084 container_of(timer, struct cfs_bandwidth, period_timer);
4088 raw_spin_lock(&cfs_b->lock);
4090 overrun = hrtimer_forward_now(timer, cfs_b->period);
4094 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4097 cfs_b->period_active = 0;
4098 raw_spin_unlock(&cfs_b->lock);
4100 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4103 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4105 raw_spin_lock_init(&cfs_b->lock);
4107 cfs_b->quota = RUNTIME_INF;
4108 cfs_b->period = ns_to_ktime(default_cfs_period());
4110 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4111 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4112 cfs_b->period_timer.function = sched_cfs_period_timer;
4113 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4114 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4117 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4119 cfs_rq->runtime_enabled = 0;
4120 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4123 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4125 lockdep_assert_held(&cfs_b->lock);
4127 if (!cfs_b->period_active) {
4128 cfs_b->period_active = 1;
4129 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4130 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4134 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4136 /* init_cfs_bandwidth() was not called */
4137 if (!cfs_b->throttled_cfs_rq.next)
4140 hrtimer_cancel(&cfs_b->period_timer);
4141 hrtimer_cancel(&cfs_b->slack_timer);
4144 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4146 struct cfs_rq *cfs_rq;
4148 for_each_leaf_cfs_rq(rq, cfs_rq) {
4149 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4151 raw_spin_lock(&cfs_b->lock);
4152 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4153 raw_spin_unlock(&cfs_b->lock);
4157 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4159 struct cfs_rq *cfs_rq;
4161 for_each_leaf_cfs_rq(rq, cfs_rq) {
4162 if (!cfs_rq->runtime_enabled)
4166 * clock_task is not advancing so we just need to make sure
4167 * there's some valid quota amount
4169 cfs_rq->runtime_remaining = 1;
4171 * Offline rq is schedulable till cpu is completely disabled
4172 * in take_cpu_down(), so we prevent new cfs throttling here.
4174 cfs_rq->runtime_enabled = 0;
4176 if (cfs_rq_throttled(cfs_rq))
4177 unthrottle_cfs_rq(cfs_rq);
4181 #else /* CONFIG_CFS_BANDWIDTH */
4182 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4184 return rq_clock_task(rq_of(cfs_rq));
4187 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4188 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4189 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4190 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4192 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4197 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4202 static inline int throttled_lb_pair(struct task_group *tg,
4203 int src_cpu, int dest_cpu)
4208 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4210 #ifdef CONFIG_FAIR_GROUP_SCHED
4211 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4214 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4218 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4219 static inline void update_runtime_enabled(struct rq *rq) {}
4220 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4222 #endif /* CONFIG_CFS_BANDWIDTH */
4224 /**************************************************
4225 * CFS operations on tasks:
4228 #ifdef CONFIG_SCHED_HRTICK
4229 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4231 struct sched_entity *se = &p->se;
4232 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4234 WARN_ON(task_rq(p) != rq);
4236 if (cfs_rq->nr_running > 1) {
4237 u64 slice = sched_slice(cfs_rq, se);
4238 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4239 s64 delta = slice - ran;
4246 hrtick_start(rq, delta);
4251 * called from enqueue/dequeue and updates the hrtick when the
4252 * current task is from our class and nr_running is low enough
4255 static void hrtick_update(struct rq *rq)
4257 struct task_struct *curr = rq->curr;
4259 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4262 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4263 hrtick_start_fair(rq, curr);
4265 #else /* !CONFIG_SCHED_HRTICK */
4267 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4271 static inline void hrtick_update(struct rq *rq)
4277 static bool cpu_overutilized(int cpu);
4278 unsigned long boosted_cpu_util(int cpu);
4280 #define boosted_cpu_util(cpu) cpu_util(cpu)
4284 static void update_capacity_of(int cpu)
4286 unsigned long req_cap;
4291 /* Convert scale-invariant capacity to cpu. */
4292 req_cap = boosted_cpu_util(cpu);
4293 req_cap = req_cap * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
4294 set_cfs_cpu_capacity(cpu, true, req_cap);
4299 * The enqueue_task method is called before nr_running is
4300 * increased. Here we update the fair scheduling stats and
4301 * then put the task into the rbtree:
4304 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4306 struct cfs_rq *cfs_rq;
4307 struct sched_entity *se = &p->se;
4309 int task_new = flags & ENQUEUE_WAKEUP_NEW;
4310 int task_wakeup = flags & ENQUEUE_WAKEUP;
4314 * If in_iowait is set, the code below may not trigger any cpufreq
4315 * utilization updates, so do it here explicitly with the IOWAIT flag
4319 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4321 for_each_sched_entity(se) {
4324 cfs_rq = cfs_rq_of(se);
4325 enqueue_entity(cfs_rq, se, flags);
4328 * end evaluation on encountering a throttled cfs_rq
4330 * note: in the case of encountering a throttled cfs_rq we will
4331 * post the final h_nr_running increment below.
4333 if (cfs_rq_throttled(cfs_rq))
4335 cfs_rq->h_nr_running++;
4336 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4338 flags = ENQUEUE_WAKEUP;
4341 for_each_sched_entity(se) {
4342 cfs_rq = cfs_rq_of(se);
4343 cfs_rq->h_nr_running++;
4344 walt_inc_cfs_cumulative_runnable_avg(cfs_rq, p);
4346 if (cfs_rq_throttled(cfs_rq))
4349 update_load_avg(se, 1);
4350 update_cfs_shares(cfs_rq);
4354 add_nr_running(rq, 1);
4359 * Update SchedTune accounting.
4361 * We do it before updating the CPU capacity to ensure the
4362 * boost value of the current task is accounted for in the
4363 * selection of the OPP.
4365 * We do it also in the case where we enqueue a throttled task;
4366 * we could argue that a throttled task should not boost a CPU,
4368 * a) properly implementing CPU boosting considering throttled
4369 * tasks will increase a lot the complexity of the solution
4370 * b) it's not easy to quantify the benefits introduced by
4371 * such a more complex solution.
4372 * Thus, for the time being we go for the simple solution and boost
4373 * also for throttled RQs.
4375 schedtune_enqueue_task(p, cpu_of(rq));
4378 walt_inc_cumulative_runnable_avg(rq, p);
4379 if (!task_new && !rq->rd->overutilized &&
4380 cpu_overutilized(rq->cpu)) {
4381 rq->rd->overutilized = true;
4382 trace_sched_overutilized(true);
4386 * We want to potentially trigger a freq switch
4387 * request only for tasks that are waking up; this is
4388 * because we get here also during load balancing, but
4389 * in these cases it seems wise to trigger as single
4390 * request after load balancing is done.
4392 if (task_new || task_wakeup)
4393 update_capacity_of(cpu_of(rq));
4396 #endif /* CONFIG_SMP */
4400 static void set_next_buddy(struct sched_entity *se);
4403 * The dequeue_task method is called before nr_running is
4404 * decreased. We remove the task from the rbtree and
4405 * update the fair scheduling stats:
4407 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4409 struct cfs_rq *cfs_rq;
4410 struct sched_entity *se = &p->se;
4411 int task_sleep = flags & DEQUEUE_SLEEP;
4413 for_each_sched_entity(se) {
4414 cfs_rq = cfs_rq_of(se);
4415 dequeue_entity(cfs_rq, se, flags);
4418 * end evaluation on encountering a throttled cfs_rq
4420 * note: in the case of encountering a throttled cfs_rq we will
4421 * post the final h_nr_running decrement below.
4423 if (cfs_rq_throttled(cfs_rq))
4425 cfs_rq->h_nr_running--;
4426 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4428 /* Don't dequeue parent if it has other entities besides us */
4429 if (cfs_rq->load.weight) {
4430 /* Avoid re-evaluating load for this entity: */
4431 se = parent_entity(se);
4433 * Bias pick_next to pick a task from this cfs_rq, as
4434 * p is sleeping when it is within its sched_slice.
4436 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4440 flags |= DEQUEUE_SLEEP;
4443 for_each_sched_entity(se) {
4444 cfs_rq = cfs_rq_of(se);
4445 cfs_rq->h_nr_running--;
4446 walt_dec_cfs_cumulative_runnable_avg(cfs_rq, p);
4448 if (cfs_rq_throttled(cfs_rq))
4451 update_load_avg(se, 1);
4452 update_cfs_shares(cfs_rq);
4456 sub_nr_running(rq, 1);
4461 * Update SchedTune accounting
4463 * We do it before updating the CPU capacity to ensure the
4464 * boost value of the current task is accounted for in the
4465 * selection of the OPP.
4467 schedtune_dequeue_task(p, cpu_of(rq));
4470 walt_dec_cumulative_runnable_avg(rq, p);
4473 * We want to potentially trigger a freq switch
4474 * request only for tasks that are going to sleep;
4475 * this is because we get here also during load
4476 * balancing, but in these cases it seems wise to
4477 * trigger as single request after load balancing is
4481 if (rq->cfs.nr_running)
4482 update_capacity_of(cpu_of(rq));
4483 else if (sched_freq())
4484 set_cfs_cpu_capacity(cpu_of(rq), false, 0);
4488 #endif /* CONFIG_SMP */
4496 * per rq 'load' arrray crap; XXX kill this.
4500 * The exact cpuload at various idx values, calculated at every tick would be
4501 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4503 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4504 * on nth tick when cpu may be busy, then we have:
4505 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4506 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4508 * decay_load_missed() below does efficient calculation of
4509 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4510 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4512 * The calculation is approximated on a 128 point scale.
4513 * degrade_zero_ticks is the number of ticks after which load at any
4514 * particular idx is approximated to be zero.
4515 * degrade_factor is a precomputed table, a row for each load idx.
4516 * Each column corresponds to degradation factor for a power of two ticks,
4517 * based on 128 point scale.
4519 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4520 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4522 * With this power of 2 load factors, we can degrade the load n times
4523 * by looking at 1 bits in n and doing as many mult/shift instead of
4524 * n mult/shifts needed by the exact degradation.
4526 #define DEGRADE_SHIFT 7
4527 static const unsigned char
4528 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4529 static const unsigned char
4530 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4531 {0, 0, 0, 0, 0, 0, 0, 0},
4532 {64, 32, 8, 0, 0, 0, 0, 0},
4533 {96, 72, 40, 12, 1, 0, 0},
4534 {112, 98, 75, 43, 15, 1, 0},
4535 {120, 112, 98, 76, 45, 16, 2} };
4538 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4539 * would be when CPU is idle and so we just decay the old load without
4540 * adding any new load.
4542 static unsigned long
4543 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4547 if (!missed_updates)
4550 if (missed_updates >= degrade_zero_ticks[idx])
4554 return load >> missed_updates;
4556 while (missed_updates) {
4557 if (missed_updates % 2)
4558 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4560 missed_updates >>= 1;
4567 * Update rq->cpu_load[] statistics. This function is usually called every
4568 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4569 * every tick. We fix it up based on jiffies.
4571 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4572 unsigned long pending_updates)
4576 this_rq->nr_load_updates++;
4578 /* Update our load: */
4579 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4580 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4581 unsigned long old_load, new_load;
4583 /* scale is effectively 1 << i now, and >> i divides by scale */
4585 old_load = this_rq->cpu_load[i];
4586 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4587 new_load = this_load;
4589 * Round up the averaging division if load is increasing. This
4590 * prevents us from getting stuck on 9 if the load is 10, for
4593 if (new_load > old_load)
4594 new_load += scale - 1;
4596 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4599 sched_avg_update(this_rq);
4602 /* Used instead of source_load when we know the type == 0 */
4603 static unsigned long weighted_cpuload(const int cpu)
4605 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4608 #ifdef CONFIG_NO_HZ_COMMON
4610 * There is no sane way to deal with nohz on smp when using jiffies because the
4611 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4612 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4614 * Therefore we cannot use the delta approach from the regular tick since that
4615 * would seriously skew the load calculation. However we'll make do for those
4616 * updates happening while idle (nohz_idle_balance) or coming out of idle
4617 * (tick_nohz_idle_exit).
4619 * This means we might still be one tick off for nohz periods.
4623 * Called from nohz_idle_balance() to update the load ratings before doing the
4626 static void update_idle_cpu_load(struct rq *this_rq)
4628 unsigned long curr_jiffies = READ_ONCE(jiffies);
4629 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4630 unsigned long pending_updates;
4633 * bail if there's load or we're actually up-to-date.
4635 if (load || curr_jiffies == this_rq->last_load_update_tick)
4638 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4639 this_rq->last_load_update_tick = curr_jiffies;
4641 __update_cpu_load(this_rq, load, pending_updates);
4645 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4647 void update_cpu_load_nohz(void)
4649 struct rq *this_rq = this_rq();
4650 unsigned long curr_jiffies = READ_ONCE(jiffies);
4651 unsigned long pending_updates;
4653 if (curr_jiffies == this_rq->last_load_update_tick)
4656 raw_spin_lock(&this_rq->lock);
4657 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4658 if (pending_updates) {
4659 this_rq->last_load_update_tick = curr_jiffies;
4661 * We were idle, this means load 0, the current load might be
4662 * !0 due to remote wakeups and the sort.
4664 __update_cpu_load(this_rq, 0, pending_updates);
4666 raw_spin_unlock(&this_rq->lock);
4668 #endif /* CONFIG_NO_HZ */
4671 * Called from scheduler_tick()
4673 void update_cpu_load_active(struct rq *this_rq)
4675 unsigned long load = weighted_cpuload(cpu_of(this_rq));
4677 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4679 this_rq->last_load_update_tick = jiffies;
4680 __update_cpu_load(this_rq, load, 1);
4684 * Return a low guess at the load of a migration-source cpu weighted
4685 * according to the scheduling class and "nice" value.
4687 * We want to under-estimate the load of migration sources, to
4688 * balance conservatively.
4690 static unsigned long source_load(int cpu, int type)
4692 struct rq *rq = cpu_rq(cpu);
4693 unsigned long total = weighted_cpuload(cpu);
4695 if (type == 0 || !sched_feat(LB_BIAS))
4698 return min(rq->cpu_load[type-1], total);
4702 * Return a high guess at the load of a migration-target cpu weighted
4703 * according to the scheduling class and "nice" value.
4705 static unsigned long target_load(int cpu, int type)
4707 struct rq *rq = cpu_rq(cpu);
4708 unsigned long total = weighted_cpuload(cpu);
4710 if (type == 0 || !sched_feat(LB_BIAS))
4713 return max(rq->cpu_load[type-1], total);
4717 static unsigned long cpu_avg_load_per_task(int cpu)
4719 struct rq *rq = cpu_rq(cpu);
4720 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4721 unsigned long load_avg = weighted_cpuload(cpu);
4724 return load_avg / nr_running;
4729 static void record_wakee(struct task_struct *p)
4732 * Rough decay (wiping) for cost saving, don't worry
4733 * about the boundary, really active task won't care
4736 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4737 current->wakee_flips >>= 1;
4738 current->wakee_flip_decay_ts = jiffies;
4741 if (current->last_wakee != p) {
4742 current->last_wakee = p;
4743 current->wakee_flips++;
4747 static void task_waking_fair(struct task_struct *p)
4749 struct sched_entity *se = &p->se;
4750 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4753 #ifndef CONFIG_64BIT
4754 u64 min_vruntime_copy;
4757 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4759 min_vruntime = cfs_rq->min_vruntime;
4760 } while (min_vruntime != min_vruntime_copy);
4762 min_vruntime = cfs_rq->min_vruntime;
4765 se->vruntime -= min_vruntime;
4769 #ifdef CONFIG_FAIR_GROUP_SCHED
4771 * effective_load() calculates the load change as seen from the root_task_group
4773 * Adding load to a group doesn't make a group heavier, but can cause movement
4774 * of group shares between cpus. Assuming the shares were perfectly aligned one
4775 * can calculate the shift in shares.
4777 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4778 * on this @cpu and results in a total addition (subtraction) of @wg to the
4779 * total group weight.
4781 * Given a runqueue weight distribution (rw_i) we can compute a shares
4782 * distribution (s_i) using:
4784 * s_i = rw_i / \Sum rw_j (1)
4786 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4787 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4788 * shares distribution (s_i):
4790 * rw_i = { 2, 4, 1, 0 }
4791 * s_i = { 2/7, 4/7, 1/7, 0 }
4793 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4794 * task used to run on and the CPU the waker is running on), we need to
4795 * compute the effect of waking a task on either CPU and, in case of a sync
4796 * wakeup, compute the effect of the current task going to sleep.
4798 * So for a change of @wl to the local @cpu with an overall group weight change
4799 * of @wl we can compute the new shares distribution (s'_i) using:
4801 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4803 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4804 * differences in waking a task to CPU 0. The additional task changes the
4805 * weight and shares distributions like:
4807 * rw'_i = { 3, 4, 1, 0 }
4808 * s'_i = { 3/8, 4/8, 1/8, 0 }
4810 * We can then compute the difference in effective weight by using:
4812 * dw_i = S * (s'_i - s_i) (3)
4814 * Where 'S' is the group weight as seen by its parent.
4816 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4817 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4818 * 4/7) times the weight of the group.
4820 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4822 struct sched_entity *se = tg->se[cpu];
4824 if (!tg->parent) /* the trivial, non-cgroup case */
4827 for_each_sched_entity(se) {
4828 struct cfs_rq *cfs_rq = se->my_q;
4829 long W, w = cfs_rq_load_avg(cfs_rq);
4834 * W = @wg + \Sum rw_j
4836 W = wg + atomic_long_read(&tg->load_avg);
4838 /* Ensure \Sum rw_j >= rw_i */
4839 W -= cfs_rq->tg_load_avg_contrib;
4848 * wl = S * s'_i; see (2)
4851 wl = (w * (long)tg->shares) / W;
4856 * Per the above, wl is the new se->load.weight value; since
4857 * those are clipped to [MIN_SHARES, ...) do so now. See
4858 * calc_cfs_shares().
4860 if (wl < MIN_SHARES)
4864 * wl = dw_i = S * (s'_i - s_i); see (3)
4866 wl -= se->avg.load_avg;
4869 * Recursively apply this logic to all parent groups to compute
4870 * the final effective load change on the root group. Since
4871 * only the @tg group gets extra weight, all parent groups can
4872 * only redistribute existing shares. @wl is the shift in shares
4873 * resulting from this level per the above.
4882 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4890 * Returns the current capacity of cpu after applying both
4891 * cpu and freq scaling.
4893 unsigned long capacity_curr_of(int cpu)
4895 return cpu_rq(cpu)->cpu_capacity_orig *
4896 arch_scale_freq_capacity(NULL, cpu)
4897 >> SCHED_CAPACITY_SHIFT;
4900 static inline bool energy_aware(void)
4902 return sched_feat(ENERGY_AWARE);
4906 struct sched_group *sg_top;
4907 struct sched_group *sg_cap;
4914 struct task_struct *task;
4929 * __cpu_norm_util() returns the cpu util relative to a specific capacity,
4930 * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
4931 * energy calculations. Using the scale-invariant util returned by
4932 * cpu_util() and approximating scale-invariant util by:
4934 * util ~ (curr_freq/max_freq)*1024 * capacity_orig/1024 * running_time/time
4936 * the normalized util can be found using the specific capacity.
4938 * capacity = capacity_orig * curr_freq/max_freq
4940 * norm_util = running_time/time ~ util/capacity
4942 static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
4944 int util = __cpu_util(cpu, delta);
4946 if (util >= capacity)
4947 return SCHED_CAPACITY_SCALE;
4949 return (util << SCHED_CAPACITY_SHIFT)/capacity;
4952 static int calc_util_delta(struct energy_env *eenv, int cpu)
4954 if (cpu == eenv->src_cpu)
4955 return -eenv->util_delta;
4956 if (cpu == eenv->dst_cpu)
4957 return eenv->util_delta;
4962 unsigned long group_max_util(struct energy_env *eenv)
4965 unsigned long max_util = 0;
4967 for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
4968 delta = calc_util_delta(eenv, i);
4969 max_util = max(max_util, __cpu_util(i, delta));
4976 * group_norm_util() returns the approximated group util relative to it's
4977 * current capacity (busy ratio) in the range [0..SCHED_LOAD_SCALE] for use in
4978 * energy calculations. Since task executions may or may not overlap in time in
4979 * the group the true normalized util is between max(cpu_norm_util(i)) and
4980 * sum(cpu_norm_util(i)) when iterating over all cpus in the group, i. The
4981 * latter is used as the estimate as it leads to a more pessimistic energy
4982 * estimate (more busy).
4985 long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
4988 unsigned long util_sum = 0;
4989 unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
4991 for_each_cpu(i, sched_group_cpus(sg)) {
4992 delta = calc_util_delta(eenv, i);
4993 util_sum += __cpu_norm_util(i, capacity, delta);
4996 if (util_sum > SCHED_CAPACITY_SCALE)
4997 return SCHED_CAPACITY_SCALE;
5001 static int find_new_capacity(struct energy_env *eenv,
5002 const struct sched_group_energy * const sge)
5005 unsigned long util = group_max_util(eenv);
5007 for (idx = 0; idx < sge->nr_cap_states; idx++) {
5008 if (sge->cap_states[idx].cap >= util)
5012 eenv->cap_idx = idx;
5017 static int group_idle_state(struct sched_group *sg)
5019 int i, state = INT_MAX;
5021 /* Find the shallowest idle state in the sched group. */
5022 for_each_cpu(i, sched_group_cpus(sg))
5023 state = min(state, idle_get_state_idx(cpu_rq(i)));
5025 /* Take non-cpuidle idling into account (active idle/arch_cpu_idle()) */
5032 * sched_group_energy(): Computes the absolute energy consumption of cpus
5033 * belonging to the sched_group including shared resources shared only by
5034 * members of the group. Iterates over all cpus in the hierarchy below the
5035 * sched_group starting from the bottom working it's way up before going to
5036 * the next cpu until all cpus are covered at all levels. The current
5037 * implementation is likely to gather the same util statistics multiple times.
5038 * This can probably be done in a faster but more complex way.
5039 * Note: sched_group_energy() may fail when racing with sched_domain updates.
5041 static int sched_group_energy(struct energy_env *eenv)
5043 struct sched_domain *sd;
5044 int cpu, total_energy = 0;
5045 struct cpumask visit_cpus;
5046 struct sched_group *sg;
5048 WARN_ON(!eenv->sg_top->sge);
5050 cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
5052 while (!cpumask_empty(&visit_cpus)) {
5053 struct sched_group *sg_shared_cap = NULL;
5055 cpu = cpumask_first(&visit_cpus);
5058 * Is the group utilization affected by cpus outside this
5061 sd = rcu_dereference(per_cpu(sd_scs, cpu));
5065 * We most probably raced with hotplug; returning a
5066 * wrong energy estimation is better than entering an
5072 sg_shared_cap = sd->parent->groups;
5074 for_each_domain(cpu, sd) {
5077 /* Has this sched_domain already been visited? */
5078 if (sd->child && group_first_cpu(sg) != cpu)
5082 unsigned long group_util;
5083 int sg_busy_energy, sg_idle_energy;
5084 int cap_idx, idle_idx;
5086 if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
5087 eenv->sg_cap = sg_shared_cap;
5091 cap_idx = find_new_capacity(eenv, sg->sge);
5093 if (sg->group_weight == 1) {
5094 /* Remove capacity of src CPU (before task move) */
5095 if (eenv->util_delta == 0 &&
5096 cpumask_test_cpu(eenv->src_cpu, sched_group_cpus(sg))) {
5097 eenv->cap.before = sg->sge->cap_states[cap_idx].cap;
5098 eenv->cap.delta -= eenv->cap.before;
5100 /* Add capacity of dst CPU (after task move) */
5101 if (eenv->util_delta != 0 &&
5102 cpumask_test_cpu(eenv->dst_cpu, sched_group_cpus(sg))) {
5103 eenv->cap.after = sg->sge->cap_states[cap_idx].cap;
5104 eenv->cap.delta += eenv->cap.after;
5108 idle_idx = group_idle_state(sg);
5109 group_util = group_norm_util(eenv, sg);
5110 sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
5111 >> SCHED_CAPACITY_SHIFT;
5112 sg_idle_energy = ((SCHED_LOAD_SCALE-group_util)
5113 * sg->sge->idle_states[idle_idx].power)
5114 >> SCHED_CAPACITY_SHIFT;
5116 total_energy += sg_busy_energy + sg_idle_energy;
5119 cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
5121 if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
5124 } while (sg = sg->next, sg != sd->groups);
5127 cpumask_clear_cpu(cpu, &visit_cpus);
5131 eenv->energy = total_energy;
5135 static inline bool cpu_in_sg(struct sched_group *sg, int cpu)
5137 return cpu != -1 && cpumask_test_cpu(cpu, sched_group_cpus(sg));
5141 * energy_diff(): Estimate the energy impact of changing the utilization
5142 * distribution. eenv specifies the change: utilisation amount, source, and
5143 * destination cpu. Source or destination cpu may be -1 in which case the
5144 * utilization is removed from or added to the system (e.g. task wake-up). If
5145 * both are specified, the utilization is migrated.
5147 static inline int __energy_diff(struct energy_env *eenv)
5149 struct sched_domain *sd;
5150 struct sched_group *sg;
5151 int sd_cpu = -1, energy_before = 0, energy_after = 0;
5154 struct energy_env eenv_before = {
5156 .src_cpu = eenv->src_cpu,
5157 .dst_cpu = eenv->dst_cpu,
5158 .nrg = { 0, 0, 0, 0},
5162 if (eenv->src_cpu == eenv->dst_cpu)
5165 sd_cpu = (eenv->src_cpu != -1) ? eenv->src_cpu : eenv->dst_cpu;
5166 sd = rcu_dereference(per_cpu(sd_ea, sd_cpu));
5169 return 0; /* Error */
5174 if (cpu_in_sg(sg, eenv->src_cpu) || cpu_in_sg(sg, eenv->dst_cpu)) {
5175 eenv_before.sg_top = eenv->sg_top = sg;
5177 if (sched_group_energy(&eenv_before))
5178 return 0; /* Invalid result abort */
5179 energy_before += eenv_before.energy;
5181 /* Keep track of SRC cpu (before) capacity */
5182 eenv->cap.before = eenv_before.cap.before;
5183 eenv->cap.delta = eenv_before.cap.delta;
5185 if (sched_group_energy(eenv))
5186 return 0; /* Invalid result abort */
5187 energy_after += eenv->energy;
5189 } while (sg = sg->next, sg != sd->groups);
5191 eenv->nrg.before = energy_before;
5192 eenv->nrg.after = energy_after;
5193 eenv->nrg.diff = eenv->nrg.after - eenv->nrg.before;
5196 trace_sched_energy_diff(eenv->task,
5197 eenv->src_cpu, eenv->dst_cpu, eenv->util_delta,
5198 eenv->nrg.before, eenv->nrg.after, eenv->nrg.diff,
5199 eenv->cap.before, eenv->cap.after, eenv->cap.delta,
5200 eenv->nrg.delta, eenv->payoff);
5203 * Dead-zone margin preventing too many migrations.
5206 margin = eenv->nrg.before >> 6; /* ~1.56% */
5208 diff = eenv->nrg.after - eenv->nrg.before;
5210 eenv->nrg.diff = (abs(diff) < margin) ? 0 : eenv->nrg.diff;
5212 return eenv->nrg.diff;
5215 #ifdef CONFIG_SCHED_TUNE
5217 struct target_nrg schedtune_target_nrg;
5220 * System energy normalization
5221 * Returns the normalized value, in the range [0..SCHED_LOAD_SCALE],
5222 * corresponding to the specified energy variation.
5225 normalize_energy(int energy_diff)
5228 #ifdef CONFIG_SCHED_DEBUG
5231 /* Check for boundaries */
5232 max_delta = schedtune_target_nrg.max_power;
5233 max_delta -= schedtune_target_nrg.min_power;
5234 WARN_ON(abs(energy_diff) >= max_delta);
5237 /* Do scaling using positive numbers to increase the range */
5238 normalized_nrg = (energy_diff < 0) ? -energy_diff : energy_diff;
5240 /* Scale by energy magnitude */
5241 normalized_nrg <<= SCHED_LOAD_SHIFT;
5243 /* Normalize on max energy for target platform */
5244 normalized_nrg = reciprocal_divide(
5245 normalized_nrg, schedtune_target_nrg.rdiv);
5247 return (energy_diff < 0) ? -normalized_nrg : normalized_nrg;
5251 energy_diff(struct energy_env *eenv)
5253 int boost = schedtune_task_boost(eenv->task);
5256 /* Conpute "absolute" energy diff */
5257 __energy_diff(eenv);
5259 /* Return energy diff when boost margin is 0 */
5261 return eenv->nrg.diff;
5263 /* Compute normalized energy diff */
5264 nrg_delta = normalize_energy(eenv->nrg.diff);
5265 eenv->nrg.delta = nrg_delta;
5267 eenv->payoff = schedtune_accept_deltas(
5273 * When SchedTune is enabled, the energy_diff() function will return
5274 * the computed energy payoff value. Since the energy_diff() return
5275 * value is expected to be negative by its callers, this evaluation
5276 * function return a negative value each time the evaluation return a
5277 * positive payoff, which is the condition for the acceptance of
5278 * a scheduling decision
5280 return -eenv->payoff;
5282 #else /* CONFIG_SCHED_TUNE */
5283 #define energy_diff(eenv) __energy_diff(eenv)
5287 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5288 * A waker of many should wake a different task than the one last awakened
5289 * at a frequency roughly N times higher than one of its wakees. In order
5290 * to determine whether we should let the load spread vs consolodating to
5291 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
5292 * partner, and a factor of lls_size higher frequency in the other. With
5293 * both conditions met, we can be relatively sure that the relationship is
5294 * non-monogamous, with partner count exceeding socket size. Waker/wakee
5295 * being client/server, worker/dispatcher, interrupt source or whatever is
5296 * irrelevant, spread criteria is apparent partner count exceeds socket size.
5298 static int wake_wide(struct task_struct *p)
5300 unsigned int master = current->wakee_flips;
5301 unsigned int slave = p->wakee_flips;
5302 int factor = this_cpu_read(sd_llc_size);
5305 swap(master, slave);
5306 if (slave < factor || master < slave * factor)
5311 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5312 int prev_cpu, int sync)
5314 s64 this_load, load;
5315 s64 this_eff_load, prev_eff_load;
5317 struct task_group *tg;
5318 unsigned long weight;
5322 this_cpu = smp_processor_id();
5323 load = source_load(prev_cpu, idx);
5324 this_load = target_load(this_cpu, idx);
5327 * If sync wakeup then subtract the (maximum possible)
5328 * effect of the currently running task from the load
5329 * of the current CPU:
5332 tg = task_group(current);
5333 weight = current->se.avg.load_avg;
5335 this_load += effective_load(tg, this_cpu, -weight, -weight);
5336 load += effective_load(tg, prev_cpu, 0, -weight);
5340 weight = p->se.avg.load_avg;
5343 * In low-load situations, where prev_cpu is idle and this_cpu is idle
5344 * due to the sync cause above having dropped this_load to 0, we'll
5345 * always have an imbalance, but there's really nothing you can do
5346 * about that, so that's good too.
5348 * Otherwise check if either cpus are near enough in load to allow this
5349 * task to be woken on this_cpu.
5351 this_eff_load = 100;
5352 this_eff_load *= capacity_of(prev_cpu);
5354 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5355 prev_eff_load *= capacity_of(this_cpu);
5357 if (this_load > 0) {
5358 this_eff_load *= this_load +
5359 effective_load(tg, this_cpu, weight, weight);
5361 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5364 balanced = this_eff_load <= prev_eff_load;
5366 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5371 schedstat_inc(sd, ttwu_move_affine);
5372 schedstat_inc(p, se.statistics.nr_wakeups_affine);
5377 static inline unsigned long task_util(struct task_struct *p)
5379 #ifdef CONFIG_SCHED_WALT
5380 if (!walt_disabled && sysctl_sched_use_walt_task_util) {
5381 unsigned long demand = p->ravg.demand;
5382 return (demand << 10) / walt_ravg_window;
5385 return p->se.avg.util_avg;
5388 static inline unsigned long boosted_task_util(struct task_struct *task);
5390 static inline bool __task_fits(struct task_struct *p, int cpu, int util)
5392 unsigned long capacity = capacity_of(cpu);
5394 util += boosted_task_util(p);
5396 return (capacity * 1024) > (util * capacity_margin);
5399 static inline bool task_fits_max(struct task_struct *p, int cpu)
5401 unsigned long capacity = capacity_of(cpu);
5402 unsigned long max_capacity = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5404 if (capacity == max_capacity)
5407 if (capacity * capacity_margin > max_capacity * 1024)
5410 return __task_fits(p, cpu, 0);
5413 static bool cpu_overutilized(int cpu)
5415 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5418 #ifdef CONFIG_SCHED_TUNE
5421 schedtune_margin(unsigned long signal, long boost)
5423 long long margin = 0;
5426 * Signal proportional compensation (SPC)
5428 * The Boost (B) value is used to compute a Margin (M) which is
5429 * proportional to the complement of the original Signal (S):
5430 * M = B * (SCHED_LOAD_SCALE - S), if B is positive
5431 * M = B * S, if B is negative
5432 * The obtained M could be used by the caller to "boost" S.
5435 margin = SCHED_LOAD_SCALE - signal;
5438 margin = -signal * boost;
5440 * Fast integer division by constant:
5441 * Constant : (C) = 100
5442 * Precision : 0.1% (P) = 0.1
5443 * Reference : C * 100 / P (R) = 100000
5446 * Shift bits : ceil(log(R,2)) (S) = 17
5447 * Mult const : round(2^S/C) (M) = 1311
5460 schedtune_cpu_margin(unsigned long util, int cpu)
5462 int boost = schedtune_cpu_boost(cpu);
5467 return schedtune_margin(util, boost);
5471 schedtune_task_margin(struct task_struct *task)
5473 int boost = schedtune_task_boost(task);
5480 util = task_util(task);
5481 margin = schedtune_margin(util, boost);
5486 #else /* CONFIG_SCHED_TUNE */
5489 schedtune_cpu_margin(unsigned long util, int cpu)
5495 schedtune_task_margin(struct task_struct *task)
5500 #endif /* CONFIG_SCHED_TUNE */
5503 boosted_cpu_util(int cpu)
5505 unsigned long util = cpu_util(cpu);
5506 long margin = schedtune_cpu_margin(util, cpu);
5508 trace_sched_boost_cpu(cpu, util, margin);
5510 return util + margin;
5513 static inline unsigned long
5514 boosted_task_util(struct task_struct *task)
5516 unsigned long util = task_util(task);
5517 long margin = schedtune_task_margin(task);
5519 trace_sched_boost_task(task, util, margin);
5521 return util + margin;
5524 static int cpu_util_wake(int cpu, struct task_struct *p);
5526 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5528 return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5532 * find_idlest_group finds and returns the least busy CPU group within the
5535 static struct sched_group *
5536 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5537 int this_cpu, int sd_flag)
5539 struct sched_group *idlest = NULL, *group = sd->groups;
5540 struct sched_group *most_spare_sg = NULL;
5541 unsigned long min_load = ULONG_MAX, this_load = 0;
5542 unsigned long most_spare = 0, this_spare = 0;
5543 int load_idx = sd->forkexec_idx;
5544 int imbalance = 100 + (sd->imbalance_pct-100)/2;
5546 if (sd_flag & SD_BALANCE_WAKE)
5547 load_idx = sd->wake_idx;
5550 unsigned long load, avg_load, spare_cap, max_spare_cap;
5554 /* Skip over this group if it has no CPUs allowed */
5555 if (!cpumask_intersects(sched_group_cpus(group),
5556 tsk_cpus_allowed(p)))
5559 local_group = cpumask_test_cpu(this_cpu,
5560 sched_group_cpus(group));
5563 * Tally up the load of all CPUs in the group and find
5564 * the group containing the CPU with most spare capacity.
5569 for_each_cpu(i, sched_group_cpus(group)) {
5570 /* Bias balancing toward cpus of our domain */
5572 load = source_load(i, load_idx);
5574 load = target_load(i, load_idx);
5578 spare_cap = capacity_spare_wake(i, p);
5580 if (spare_cap > max_spare_cap)
5581 max_spare_cap = spare_cap;
5584 /* Adjust by relative CPU capacity of the group */
5585 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5588 this_load = avg_load;
5589 this_spare = max_spare_cap;
5591 if (avg_load < min_load) {
5592 min_load = avg_load;
5596 if (most_spare < max_spare_cap) {
5597 most_spare = max_spare_cap;
5598 most_spare_sg = group;
5601 } while (group = group->next, group != sd->groups);
5604 * The cross-over point between using spare capacity or least load
5605 * is too conservative for high utilization tasks on partially
5606 * utilized systems if we require spare_capacity > task_util(p),
5607 * so we allow for some task stuffing by using
5608 * spare_capacity > task_util(p)/2.
5610 if (this_spare > task_util(p) / 2 &&
5611 imbalance*this_spare > 100*most_spare)
5613 else if (most_spare > task_util(p) / 2)
5614 return most_spare_sg;
5616 if (!idlest || 100*this_load < imbalance*min_load)
5622 * find_idlest_cpu - find the idlest cpu among the cpus in group.
5625 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5627 unsigned long load, min_load = ULONG_MAX;
5628 unsigned int min_exit_latency = UINT_MAX;
5629 u64 latest_idle_timestamp = 0;
5630 int least_loaded_cpu = this_cpu;
5631 int shallowest_idle_cpu = -1;
5634 /* Check if we have any choice: */
5635 if (group->group_weight == 1)
5636 return cpumask_first(sched_group_cpus(group));
5638 /* Traverse only the allowed CPUs */
5639 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5641 struct rq *rq = cpu_rq(i);
5642 struct cpuidle_state *idle = idle_get_state(rq);
5643 if (idle && idle->exit_latency < min_exit_latency) {
5645 * We give priority to a CPU whose idle state
5646 * has the smallest exit latency irrespective
5647 * of any idle timestamp.
5649 min_exit_latency = idle->exit_latency;
5650 latest_idle_timestamp = rq->idle_stamp;
5651 shallowest_idle_cpu = i;
5652 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5653 rq->idle_stamp > latest_idle_timestamp) {
5655 * If equal or no active idle state, then
5656 * the most recently idled CPU might have
5659 latest_idle_timestamp = rq->idle_stamp;
5660 shallowest_idle_cpu = i;
5662 } else if (shallowest_idle_cpu == -1) {
5663 load = weighted_cpuload(i);
5664 if (load < min_load || (load == min_load && i == this_cpu)) {
5666 least_loaded_cpu = i;
5671 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5675 * Try and locate an idle CPU in the sched_domain.
5677 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5679 struct sched_domain *sd;
5680 struct sched_group *sg;
5682 int best_idle_cstate = -1;
5683 int best_idle_capacity = INT_MAX;
5685 if (!sysctl_sched_cstate_aware) {
5686 if (idle_cpu(target))
5690 * If the prevous cpu is cache affine and idle, don't be stupid.
5692 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5697 * Otherwise, iterate the domains and find an elegible idle cpu.
5699 sd = rcu_dereference(per_cpu(sd_llc, target));
5700 for_each_lower_domain(sd) {
5704 if (!cpumask_intersects(sched_group_cpus(sg),
5705 tsk_cpus_allowed(p)))
5708 if (sysctl_sched_cstate_aware) {
5709 for_each_cpu_and(i, tsk_cpus_allowed(p), sched_group_cpus(sg)) {
5710 struct rq *rq = cpu_rq(i);
5711 int idle_idx = idle_get_state_idx(rq);
5712 unsigned long new_usage = boosted_task_util(p);
5713 unsigned long capacity_orig = capacity_orig_of(i);
5714 if (new_usage > capacity_orig || !idle_cpu(i))
5717 if (i == target && new_usage <= capacity_curr_of(target))
5720 if (best_idle < 0 || (idle_idx < best_idle_cstate && 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);
5747 static inline int find_best_target(struct task_struct *p, bool boosted, bool prefer_idle)
5750 int target_cpu = -1;
5751 int target_util = 0;
5752 int backup_capacity = 0;
5753 int best_idle_cpu = -1;
5754 int best_idle_cstate = INT_MAX;
5755 int backup_cpu = -1;
5756 unsigned long task_util_boosted, new_util;
5758 task_util_boosted = boosted_task_util(p);
5759 for (iter_cpu = 0; iter_cpu < NR_CPUS; iter_cpu++) {
5765 * Iterate from higher cpus for boosted tasks.
5767 int i = boosted ? NR_CPUS-iter_cpu-1 : iter_cpu;
5769 if (!cpu_online(i) || !cpumask_test_cpu(i, tsk_cpus_allowed(p)))
5773 * p's blocked utilization is still accounted for on prev_cpu
5774 * so prev_cpu will receive a negative bias due to the double
5775 * accounting. However, the blocked utilization may be zero.
5777 new_util = cpu_util(i) + task_util_boosted;
5780 * Ensure minimum capacity to grant the required boost.
5781 * The target CPU can be already at a capacity level higher
5782 * than the one required to boost the task.
5784 if (new_util > capacity_orig_of(i))
5787 #ifdef CONFIG_SCHED_WALT
5788 if (walt_cpu_high_irqload(i))
5792 * Unconditionally favoring tasks that prefer idle cpus to
5795 if (idle_cpu(i) && prefer_idle) {
5796 if (best_idle_cpu < 0)
5801 cur_capacity = capacity_curr_of(i);
5803 idle_idx = idle_get_state_idx(rq);
5805 if (new_util < cur_capacity) {
5806 if (cpu_rq(i)->nr_running) {
5808 /* Find a target cpu with highest
5811 if (target_util == 0 ||
5812 target_util < new_util) {
5814 target_util = new_util;
5817 /* Find a target cpu with lowest
5820 if (target_util == 0 ||
5821 target_util > new_util) {
5823 target_util = new_util;
5826 } else if (!prefer_idle) {
5827 if (best_idle_cpu < 0 ||
5828 (sysctl_sched_cstate_aware &&
5829 best_idle_cstate > idle_idx)) {
5830 best_idle_cstate = idle_idx;
5834 } else if (backup_capacity == 0 ||
5835 backup_capacity > cur_capacity) {
5836 // Find a backup cpu with least capacity.
5837 backup_capacity = cur_capacity;
5842 if (prefer_idle && best_idle_cpu >= 0)
5843 target_cpu = best_idle_cpu;
5844 else if (target_cpu < 0)
5845 target_cpu = best_idle_cpu >= 0 ? best_idle_cpu : backup_cpu;
5851 * cpu_util_wake: Compute cpu utilization with any contributions from
5852 * the waking task p removed.
5854 static int cpu_util_wake(int cpu, struct task_struct *p)
5856 unsigned long util, capacity;
5858 /* Task has no contribution or is new */
5859 if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5860 return cpu_util(cpu);
5862 capacity = capacity_orig_of(cpu);
5863 util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5865 return (util >= capacity) ? capacity : util;
5869 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5870 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5872 * In that case WAKE_AFFINE doesn't make sense and we'll let
5873 * BALANCE_WAKE sort things out.
5875 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5877 long min_cap, max_cap;
5879 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5880 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity.val;
5882 /* Minimum capacity is close to max, no need to abort wake_affine */
5883 if (max_cap - min_cap < max_cap >> 3)
5886 /* Bring task utilization in sync with prev_cpu */
5887 sync_entity_load_avg(&p->se);
5889 return min_cap * 1024 < task_util(p) * capacity_margin;
5893 * select_task_rq_fair: Select target runqueue for the waking task in domains
5894 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5895 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5897 * Balances load by selecting the idlest cpu in the idlest group, or under
5898 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5900 * Returns the target cpu number.
5902 * preempt must be disabled.
5905 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5907 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5908 int cpu = smp_processor_id();
5909 int new_cpu = prev_cpu;
5910 int want_affine = 0;
5911 int sync = wake_flags & WF_SYNC;
5913 if (sd_flag & SD_BALANCE_WAKE)
5914 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5915 && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5918 for_each_domain(cpu, tmp) {
5919 if (!(tmp->flags & SD_LOAD_BALANCE))
5923 * If both cpu and prev_cpu are part of this domain,
5924 * cpu is a valid SD_WAKE_AFFINE target.
5926 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5927 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5932 if (tmp->flags & sd_flag)
5934 else if (!want_affine)
5939 sd = NULL; /* Prefer wake_affine over balance flags */
5940 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5945 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5946 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5949 struct sched_group *group;
5952 if (!(sd->flags & sd_flag)) {
5957 group = find_idlest_group(sd, p, cpu, sd_flag);
5963 new_cpu = find_idlest_cpu(group, p, cpu);
5964 if (new_cpu == -1 || new_cpu == cpu) {
5965 /* Now try balancing at a lower domain level of cpu */
5970 /* Now try balancing at a lower domain level of new_cpu */
5972 weight = sd->span_weight;
5974 for_each_domain(cpu, tmp) {
5975 if (weight <= tmp->span_weight)
5977 if (tmp->flags & sd_flag)
5980 /* while loop will break here if sd == NULL */
5988 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5989 * cfs_rq_of(p) references at time of call are still valid and identify the
5990 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
5991 * other assumptions, including the state of rq->lock, should be made.
5993 static void migrate_task_rq_fair(struct task_struct *p)
5996 * We are supposed to update the task to "current" time, then its up to date
5997 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5998 * what current time is, so simply throw away the out-of-date time. This
5999 * will result in the wakee task is less decayed, but giving the wakee more
6000 * load sounds not bad.
6002 remove_entity_load_avg(&p->se);
6004 /* Tell new CPU we are migrated */
6005 p->se.avg.last_update_time = 0;
6007 /* We have migrated, no longer consider this task hot */
6008 p->se.exec_start = 0;
6011 static void task_dead_fair(struct task_struct *p)
6013 remove_entity_load_avg(&p->se);
6016 #define task_fits_max(p, cpu) true
6017 #endif /* CONFIG_SMP */
6019 static unsigned long
6020 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6022 unsigned long gran = sysctl_sched_wakeup_granularity;
6025 * Since its curr running now, convert the gran from real-time
6026 * to virtual-time in his units.
6028 * By using 'se' instead of 'curr' we penalize light tasks, so
6029 * they get preempted easier. That is, if 'se' < 'curr' then
6030 * the resulting gran will be larger, therefore penalizing the
6031 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6032 * be smaller, again penalizing the lighter task.
6034 * This is especially important for buddies when the leftmost
6035 * task is higher priority than the buddy.
6037 return calc_delta_fair(gran, se);
6041 * Should 'se' preempt 'curr'.
6055 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6057 s64 gran, vdiff = curr->vruntime - se->vruntime;
6062 gran = wakeup_gran(curr, se);
6069 static void set_last_buddy(struct sched_entity *se)
6071 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6074 for_each_sched_entity(se)
6075 cfs_rq_of(se)->last = se;
6078 static void set_next_buddy(struct sched_entity *se)
6080 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6083 for_each_sched_entity(se)
6084 cfs_rq_of(se)->next = se;
6087 static void set_skip_buddy(struct sched_entity *se)
6089 for_each_sched_entity(se)
6090 cfs_rq_of(se)->skip = se;
6094 * Preempt the current task with a newly woken task if needed:
6096 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6098 struct task_struct *curr = rq->curr;
6099 struct sched_entity *se = &curr->se, *pse = &p->se;
6100 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6101 int scale = cfs_rq->nr_running >= sched_nr_latency;
6102 int next_buddy_marked = 0;
6104 if (unlikely(se == pse))
6108 * This is possible from callers such as attach_tasks(), in which we
6109 * unconditionally check_prempt_curr() after an enqueue (which may have
6110 * lead to a throttle). This both saves work and prevents false
6111 * next-buddy nomination below.
6113 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6116 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6117 set_next_buddy(pse);
6118 next_buddy_marked = 1;
6122 * We can come here with TIF_NEED_RESCHED already set from new task
6125 * Note: this also catches the edge-case of curr being in a throttled
6126 * group (e.g. via set_curr_task), since update_curr() (in the
6127 * enqueue of curr) will have resulted in resched being set. This
6128 * prevents us from potentially nominating it as a false LAST_BUDDY
6131 if (test_tsk_need_resched(curr))
6134 /* Idle tasks are by definition preempted by non-idle tasks. */
6135 if (unlikely(curr->policy == SCHED_IDLE) &&
6136 likely(p->policy != SCHED_IDLE))
6140 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6141 * is driven by the tick):
6143 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6146 find_matching_se(&se, &pse);
6147 update_curr(cfs_rq_of(se));
6149 if (wakeup_preempt_entity(se, pse) == 1) {
6151 * Bias pick_next to pick the sched entity that is
6152 * triggering this preemption.
6154 if (!next_buddy_marked)
6155 set_next_buddy(pse);
6164 * Only set the backward buddy when the current task is still
6165 * on the rq. This can happen when a wakeup gets interleaved
6166 * with schedule on the ->pre_schedule() or idle_balance()
6167 * point, either of which can * drop the rq lock.
6169 * Also, during early boot the idle thread is in the fair class,
6170 * for obvious reasons its a bad idea to schedule back to it.
6172 if (unlikely(!se->on_rq || curr == rq->idle))
6175 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6179 static struct task_struct *
6180 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
6182 struct cfs_rq *cfs_rq = &rq->cfs;
6183 struct sched_entity *se;
6184 struct task_struct *p;
6188 #ifdef CONFIG_FAIR_GROUP_SCHED
6189 if (!cfs_rq->nr_running)
6192 if (prev->sched_class != &fair_sched_class)
6196 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6197 * likely that a next task is from the same cgroup as the current.
6199 * Therefore attempt to avoid putting and setting the entire cgroup
6200 * hierarchy, only change the part that actually changes.
6204 struct sched_entity *curr = cfs_rq->curr;
6207 * Since we got here without doing put_prev_entity() we also
6208 * have to consider cfs_rq->curr. If it is still a runnable
6209 * entity, update_curr() will update its vruntime, otherwise
6210 * forget we've ever seen it.
6214 update_curr(cfs_rq);
6219 * This call to check_cfs_rq_runtime() will do the
6220 * throttle and dequeue its entity in the parent(s).
6221 * Therefore the 'simple' nr_running test will indeed
6224 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6228 se = pick_next_entity(cfs_rq, curr);
6229 cfs_rq = group_cfs_rq(se);
6235 * Since we haven't yet done put_prev_entity and if the selected task
6236 * is a different task than we started out with, try and touch the
6237 * least amount of cfs_rqs.
6240 struct sched_entity *pse = &prev->se;
6242 while (!(cfs_rq = is_same_group(se, pse))) {
6243 int se_depth = se->depth;
6244 int pse_depth = pse->depth;
6246 if (se_depth <= pse_depth) {
6247 put_prev_entity(cfs_rq_of(pse), pse);
6248 pse = parent_entity(pse);
6250 if (se_depth >= pse_depth) {
6251 set_next_entity(cfs_rq_of(se), se);
6252 se = parent_entity(se);
6256 put_prev_entity(cfs_rq, pse);
6257 set_next_entity(cfs_rq, se);
6260 if (hrtick_enabled(rq))
6261 hrtick_start_fair(rq, p);
6263 rq->misfit_task = !task_fits_max(p, rq->cpu);
6270 if (!cfs_rq->nr_running)
6273 put_prev_task(rq, prev);
6276 se = pick_next_entity(cfs_rq, NULL);
6277 set_next_entity(cfs_rq, se);
6278 cfs_rq = group_cfs_rq(se);
6283 if (hrtick_enabled(rq))
6284 hrtick_start_fair(rq, p);
6286 rq->misfit_task = !task_fits_max(p, rq->cpu);
6291 rq->misfit_task = 0;
6293 * This is OK, because current is on_cpu, which avoids it being picked
6294 * for load-balance and preemption/IRQs are still disabled avoiding
6295 * further scheduler activity on it and we're being very careful to
6296 * re-start the picking loop.
6298 lockdep_unpin_lock(&rq->lock);
6299 new_tasks = idle_balance(rq);
6300 lockdep_pin_lock(&rq->lock);
6302 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6303 * possible for any higher priority task to appear. In that case we
6304 * must re-start the pick_next_entity() loop.
6316 * Account for a descheduled task:
6318 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6320 struct sched_entity *se = &prev->se;
6321 struct cfs_rq *cfs_rq;
6323 for_each_sched_entity(se) {
6324 cfs_rq = cfs_rq_of(se);
6325 put_prev_entity(cfs_rq, se);
6330 * sched_yield() is very simple
6332 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6334 static void yield_task_fair(struct rq *rq)
6336 struct task_struct *curr = rq->curr;
6337 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6338 struct sched_entity *se = &curr->se;
6341 * Are we the only task in the tree?
6343 if (unlikely(rq->nr_running == 1))
6346 clear_buddies(cfs_rq, se);
6348 if (curr->policy != SCHED_BATCH) {
6349 update_rq_clock(rq);
6351 * Update run-time statistics of the 'current'.
6353 update_curr(cfs_rq);
6355 * Tell update_rq_clock() that we've just updated,
6356 * so we don't do microscopic update in schedule()
6357 * and double the fastpath cost.
6359 rq_clock_skip_update(rq, true);
6365 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6367 struct sched_entity *se = &p->se;
6369 /* throttled hierarchies are not runnable */
6370 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6373 /* Tell the scheduler that we'd really like pse to run next. */
6376 yield_task_fair(rq);
6382 /**************************************************
6383 * Fair scheduling class load-balancing methods.
6387 * The purpose of load-balancing is to achieve the same basic fairness the
6388 * per-cpu scheduler provides, namely provide a proportional amount of compute
6389 * time to each task. This is expressed in the following equation:
6391 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6393 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6394 * W_i,0 is defined as:
6396 * W_i,0 = \Sum_j w_i,j (2)
6398 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6399 * is derived from the nice value as per prio_to_weight[].
6401 * The weight average is an exponential decay average of the instantaneous
6404 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6406 * C_i is the compute capacity of cpu i, typically it is the
6407 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6408 * can also include other factors [XXX].
6410 * To achieve this balance we define a measure of imbalance which follows
6411 * directly from (1):
6413 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6415 * We them move tasks around to minimize the imbalance. In the continuous
6416 * function space it is obvious this converges, in the discrete case we get
6417 * a few fun cases generally called infeasible weight scenarios.
6420 * - infeasible weights;
6421 * - local vs global optima in the discrete case. ]
6426 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6427 * for all i,j solution, we create a tree of cpus that follows the hardware
6428 * topology where each level pairs two lower groups (or better). This results
6429 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6430 * tree to only the first of the previous level and we decrease the frequency
6431 * of load-balance at each level inv. proportional to the number of cpus in
6437 * \Sum { --- * --- * 2^i } = O(n) (5)
6439 * `- size of each group
6440 * | | `- number of cpus doing load-balance
6442 * `- sum over all levels
6444 * Coupled with a limit on how many tasks we can migrate every balance pass,
6445 * this makes (5) the runtime complexity of the balancer.
6447 * An important property here is that each CPU is still (indirectly) connected
6448 * to every other cpu in at most O(log n) steps:
6450 * The adjacency matrix of the resulting graph is given by:
6453 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6456 * And you'll find that:
6458 * A^(log_2 n)_i,j != 0 for all i,j (7)
6460 * Showing there's indeed a path between every cpu in at most O(log n) steps.
6461 * The task movement gives a factor of O(m), giving a convergence complexity
6464 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6469 * In order to avoid CPUs going idle while there's still work to do, new idle
6470 * balancing is more aggressive and has the newly idle cpu iterate up the domain
6471 * tree itself instead of relying on other CPUs to bring it work.
6473 * This adds some complexity to both (5) and (8) but it reduces the total idle
6481 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6484 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6489 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6491 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6493 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6496 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6497 * rewrite all of this once again.]
6500 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6502 enum fbq_type { regular, remote, all };
6511 #define LBF_ALL_PINNED 0x01
6512 #define LBF_NEED_BREAK 0x02
6513 #define LBF_DST_PINNED 0x04
6514 #define LBF_SOME_PINNED 0x08
6517 struct sched_domain *sd;
6525 struct cpumask *dst_grpmask;
6527 enum cpu_idle_type idle;
6529 unsigned int src_grp_nr_running;
6530 /* The set of CPUs under consideration for load-balancing */
6531 struct cpumask *cpus;
6536 unsigned int loop_break;
6537 unsigned int loop_max;
6539 enum fbq_type fbq_type;
6540 enum group_type busiest_group_type;
6541 struct list_head tasks;
6545 * Is this task likely cache-hot:
6547 static int task_hot(struct task_struct *p, struct lb_env *env)
6551 lockdep_assert_held(&env->src_rq->lock);
6553 if (p->sched_class != &fair_sched_class)
6556 if (unlikely(p->policy == SCHED_IDLE))
6560 * Buddy candidates are cache hot:
6562 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6563 (&p->se == cfs_rq_of(&p->se)->next ||
6564 &p->se == cfs_rq_of(&p->se)->last))
6567 if (sysctl_sched_migration_cost == -1)
6569 if (sysctl_sched_migration_cost == 0)
6572 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6574 return delta < (s64)sysctl_sched_migration_cost;
6577 #ifdef CONFIG_NUMA_BALANCING
6579 * Returns 1, if task migration degrades locality
6580 * Returns 0, if task migration improves locality i.e migration preferred.
6581 * Returns -1, if task migration is not affected by locality.
6583 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6585 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6586 unsigned long src_faults, dst_faults;
6587 int src_nid, dst_nid;
6589 if (!static_branch_likely(&sched_numa_balancing))
6592 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6595 src_nid = cpu_to_node(env->src_cpu);
6596 dst_nid = cpu_to_node(env->dst_cpu);
6598 if (src_nid == dst_nid)
6601 /* Migrating away from the preferred node is always bad. */
6602 if (src_nid == p->numa_preferred_nid) {
6603 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6609 /* Encourage migration to the preferred node. */
6610 if (dst_nid == p->numa_preferred_nid)
6614 src_faults = group_faults(p, src_nid);
6615 dst_faults = group_faults(p, dst_nid);
6617 src_faults = task_faults(p, src_nid);
6618 dst_faults = task_faults(p, dst_nid);
6621 return dst_faults < src_faults;
6625 static inline int migrate_degrades_locality(struct task_struct *p,
6633 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6636 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6640 lockdep_assert_held(&env->src_rq->lock);
6643 * We do not migrate tasks that are:
6644 * 1) throttled_lb_pair, or
6645 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6646 * 3) running (obviously), or
6647 * 4) are cache-hot on their current CPU.
6649 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6652 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6655 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6657 env->flags |= LBF_SOME_PINNED;
6660 * Remember if this task can be migrated to any other cpu in
6661 * our sched_group. We may want to revisit it if we couldn't
6662 * meet load balance goals by pulling other tasks on src_cpu.
6664 * Also avoid computing new_dst_cpu if we have already computed
6665 * one in current iteration.
6667 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6670 /* Prevent to re-select dst_cpu via env's cpus */
6671 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6672 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6673 env->flags |= LBF_DST_PINNED;
6674 env->new_dst_cpu = cpu;
6682 /* Record that we found atleast one task that could run on dst_cpu */
6683 env->flags &= ~LBF_ALL_PINNED;
6685 if (task_running(env->src_rq, p)) {
6686 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6691 * Aggressive migration if:
6692 * 1) destination numa is preferred
6693 * 2) task is cache cold, or
6694 * 3) too many balance attempts have failed.
6696 tsk_cache_hot = migrate_degrades_locality(p, env);
6697 if (tsk_cache_hot == -1)
6698 tsk_cache_hot = task_hot(p, env);
6700 if (tsk_cache_hot <= 0 ||
6701 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6702 if (tsk_cache_hot == 1) {
6703 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6704 schedstat_inc(p, se.statistics.nr_forced_migrations);
6709 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6714 * detach_task() -- detach the task for the migration specified in env
6716 static void detach_task(struct task_struct *p, struct lb_env *env)
6718 lockdep_assert_held(&env->src_rq->lock);
6720 deactivate_task(env->src_rq, p, 0);
6721 p->on_rq = TASK_ON_RQ_MIGRATING;
6722 double_lock_balance(env->src_rq, env->dst_rq);
6723 set_task_cpu(p, env->dst_cpu);
6724 double_unlock_balance(env->src_rq, env->dst_rq);
6728 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6729 * part of active balancing operations within "domain".
6731 * Returns a task if successful and NULL otherwise.
6733 static struct task_struct *detach_one_task(struct lb_env *env)
6735 struct task_struct *p, *n;
6737 lockdep_assert_held(&env->src_rq->lock);
6739 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6740 if (!can_migrate_task(p, env))
6743 detach_task(p, env);
6746 * Right now, this is only the second place where
6747 * lb_gained[env->idle] is updated (other is detach_tasks)
6748 * so we can safely collect stats here rather than
6749 * inside detach_tasks().
6751 schedstat_inc(env->sd, lb_gained[env->idle]);
6757 static const unsigned int sched_nr_migrate_break = 32;
6760 * detach_tasks() -- tries to detach up to imbalance weighted load from
6761 * busiest_rq, as part of a balancing operation within domain "sd".
6763 * Returns number of detached tasks if successful and 0 otherwise.
6765 static int detach_tasks(struct lb_env *env)
6767 struct list_head *tasks = &env->src_rq->cfs_tasks;
6768 struct task_struct *p;
6772 lockdep_assert_held(&env->src_rq->lock);
6774 if (env->imbalance <= 0)
6777 while (!list_empty(tasks)) {
6779 * We don't want to steal all, otherwise we may be treated likewise,
6780 * which could at worst lead to a livelock crash.
6782 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6785 p = list_first_entry(tasks, struct task_struct, se.group_node);
6788 /* We've more or less seen every task there is, call it quits */
6789 if (env->loop > env->loop_max)
6792 /* take a breather every nr_migrate tasks */
6793 if (env->loop > env->loop_break) {
6794 env->loop_break += sched_nr_migrate_break;
6795 env->flags |= LBF_NEED_BREAK;
6799 if (!can_migrate_task(p, env))
6802 load = task_h_load(p);
6804 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6807 if ((load / 2) > env->imbalance)
6810 detach_task(p, env);
6811 list_add(&p->se.group_node, &env->tasks);
6814 env->imbalance -= load;
6816 #ifdef CONFIG_PREEMPT
6818 * NEWIDLE balancing is a source of latency, so preemptible
6819 * kernels will stop after the first task is detached to minimize
6820 * the critical section.
6822 if (env->idle == CPU_NEWLY_IDLE)
6827 * We only want to steal up to the prescribed amount of
6830 if (env->imbalance <= 0)
6835 list_move_tail(&p->se.group_node, tasks);
6839 * Right now, this is one of only two places we collect this stat
6840 * so we can safely collect detach_one_task() stats here rather
6841 * than inside detach_one_task().
6843 schedstat_add(env->sd, lb_gained[env->idle], detached);
6849 * attach_task() -- attach the task detached by detach_task() to its new rq.
6851 static void attach_task(struct rq *rq, struct task_struct *p)
6853 lockdep_assert_held(&rq->lock);
6855 BUG_ON(task_rq(p) != rq);
6856 p->on_rq = TASK_ON_RQ_QUEUED;
6857 activate_task(rq, p, 0);
6858 check_preempt_curr(rq, p, 0);
6862 * attach_one_task() -- attaches the task returned from detach_one_task() to
6865 static void attach_one_task(struct rq *rq, struct task_struct *p)
6867 raw_spin_lock(&rq->lock);
6870 * We want to potentially raise target_cpu's OPP.
6872 update_capacity_of(cpu_of(rq));
6873 raw_spin_unlock(&rq->lock);
6877 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6880 static void attach_tasks(struct lb_env *env)
6882 struct list_head *tasks = &env->tasks;
6883 struct task_struct *p;
6885 raw_spin_lock(&env->dst_rq->lock);
6887 while (!list_empty(tasks)) {
6888 p = list_first_entry(tasks, struct task_struct, se.group_node);
6889 list_del_init(&p->se.group_node);
6891 attach_task(env->dst_rq, p);
6895 * We want to potentially raise env.dst_cpu's OPP.
6897 update_capacity_of(env->dst_cpu);
6899 raw_spin_unlock(&env->dst_rq->lock);
6902 #ifdef CONFIG_FAIR_GROUP_SCHED
6903 static void update_blocked_averages(int cpu)
6905 struct rq *rq = cpu_rq(cpu);
6906 struct cfs_rq *cfs_rq;
6907 unsigned long flags;
6909 raw_spin_lock_irqsave(&rq->lock, flags);
6910 update_rq_clock(rq);
6913 * Iterates the task_group tree in a bottom up fashion, see
6914 * list_add_leaf_cfs_rq() for details.
6916 for_each_leaf_cfs_rq(rq, cfs_rq) {
6917 /* throttled entities do not contribute to load */
6918 if (throttled_hierarchy(cfs_rq))
6921 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq,
6923 update_tg_load_avg(cfs_rq, 0);
6925 raw_spin_unlock_irqrestore(&rq->lock, flags);
6929 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6930 * This needs to be done in a top-down fashion because the load of a child
6931 * group is a fraction of its parents load.
6933 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6935 struct rq *rq = rq_of(cfs_rq);
6936 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6937 unsigned long now = jiffies;
6940 if (cfs_rq->last_h_load_update == now)
6943 cfs_rq->h_load_next = NULL;
6944 for_each_sched_entity(se) {
6945 cfs_rq = cfs_rq_of(se);
6946 cfs_rq->h_load_next = se;
6947 if (cfs_rq->last_h_load_update == now)
6952 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6953 cfs_rq->last_h_load_update = now;
6956 while ((se = cfs_rq->h_load_next) != NULL) {
6957 load = cfs_rq->h_load;
6958 load = div64_ul(load * se->avg.load_avg,
6959 cfs_rq_load_avg(cfs_rq) + 1);
6960 cfs_rq = group_cfs_rq(se);
6961 cfs_rq->h_load = load;
6962 cfs_rq->last_h_load_update = now;
6966 static unsigned long task_h_load(struct task_struct *p)
6968 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6970 update_cfs_rq_h_load(cfs_rq);
6971 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6972 cfs_rq_load_avg(cfs_rq) + 1);
6975 static inline void update_blocked_averages(int cpu)
6977 struct rq *rq = cpu_rq(cpu);
6978 struct cfs_rq *cfs_rq = &rq->cfs;
6979 unsigned long flags;
6981 raw_spin_lock_irqsave(&rq->lock, flags);
6982 update_rq_clock(rq);
6983 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6984 raw_spin_unlock_irqrestore(&rq->lock, flags);
6987 static unsigned long task_h_load(struct task_struct *p)
6989 return p->se.avg.load_avg;
6993 /********** Helpers for find_busiest_group ************************/
6996 * sg_lb_stats - stats of a sched_group required for load_balancing
6998 struct sg_lb_stats {
6999 unsigned long avg_load; /*Avg load across the CPUs of the group */
7000 unsigned long group_load; /* Total load over the CPUs of the group */
7001 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7002 unsigned long load_per_task;
7003 unsigned long group_capacity;
7004 unsigned long group_util; /* Total utilization of the group */
7005 unsigned int sum_nr_running; /* Nr tasks running in the group */
7006 unsigned int idle_cpus;
7007 unsigned int group_weight;
7008 enum group_type group_type;
7009 int group_no_capacity;
7010 int group_misfit_task; /* A cpu has a task too big for its capacity */
7011 #ifdef CONFIG_NUMA_BALANCING
7012 unsigned int nr_numa_running;
7013 unsigned int nr_preferred_running;
7018 * sd_lb_stats - Structure to store the statistics of a sched_domain
7019 * during load balancing.
7021 struct sd_lb_stats {
7022 struct sched_group *busiest; /* Busiest group in this sd */
7023 struct sched_group *local; /* Local group in this sd */
7024 unsigned long total_load; /* Total load of all groups in sd */
7025 unsigned long total_capacity; /* Total capacity of all groups in sd */
7026 unsigned long avg_load; /* Average load across all groups in sd */
7028 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7029 struct sg_lb_stats local_stat; /* Statistics of the local group */
7032 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7035 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7036 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7037 * We must however clear busiest_stat::avg_load because
7038 * update_sd_pick_busiest() reads this before assignment.
7040 *sds = (struct sd_lb_stats){
7044 .total_capacity = 0UL,
7047 .sum_nr_running = 0,
7048 .group_type = group_other,
7054 * get_sd_load_idx - Obtain the load index for a given sched domain.
7055 * @sd: The sched_domain whose load_idx is to be obtained.
7056 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7058 * Return: The load index.
7060 static inline int get_sd_load_idx(struct sched_domain *sd,
7061 enum cpu_idle_type idle)
7067 load_idx = sd->busy_idx;
7070 case CPU_NEWLY_IDLE:
7071 load_idx = sd->newidle_idx;
7074 load_idx = sd->idle_idx;
7081 static unsigned long scale_rt_capacity(int cpu)
7083 struct rq *rq = cpu_rq(cpu);
7084 u64 total, used, age_stamp, avg;
7088 * Since we're reading these variables without serialization make sure
7089 * we read them once before doing sanity checks on them.
7091 age_stamp = READ_ONCE(rq->age_stamp);
7092 avg = READ_ONCE(rq->rt_avg);
7093 delta = __rq_clock_broken(rq) - age_stamp;
7095 if (unlikely(delta < 0))
7098 total = sched_avg_period() + delta;
7100 used = div_u64(avg, total);
7103 * deadline bandwidth is defined at system level so we must
7104 * weight this bandwidth with the max capacity of the system.
7105 * As a reminder, avg_bw is 20bits width and
7106 * scale_cpu_capacity is 10 bits width
7108 used += div_u64(rq->dl.avg_bw, arch_scale_cpu_capacity(NULL, cpu));
7110 if (likely(used < SCHED_CAPACITY_SCALE))
7111 return SCHED_CAPACITY_SCALE - used;
7116 void init_max_cpu_capacity(struct max_cpu_capacity *mcc)
7118 raw_spin_lock_init(&mcc->lock);
7123 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7125 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7126 struct sched_group *sdg = sd->groups;
7127 struct max_cpu_capacity *mcc;
7128 unsigned long max_capacity;
7130 unsigned long flags;
7132 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7134 mcc = &cpu_rq(cpu)->rd->max_cpu_capacity;
7136 raw_spin_lock_irqsave(&mcc->lock, flags);
7137 max_capacity = mcc->val;
7138 max_cap_cpu = mcc->cpu;
7140 if ((max_capacity > capacity && max_cap_cpu == cpu) ||
7141 (max_capacity < capacity)) {
7142 mcc->val = capacity;
7144 #ifdef CONFIG_SCHED_DEBUG
7145 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7146 printk_deferred(KERN_INFO "CPU%d: update max cpu_capacity %lu\n",
7151 raw_spin_unlock_irqrestore(&mcc->lock, flags);
7153 skip_unlock: __attribute__ ((unused));
7154 capacity *= scale_rt_capacity(cpu);
7155 capacity >>= SCHED_CAPACITY_SHIFT;
7160 cpu_rq(cpu)->cpu_capacity = capacity;
7161 sdg->sgc->capacity = capacity;
7162 sdg->sgc->max_capacity = capacity;
7163 sdg->sgc->min_capacity = capacity;
7166 void update_group_capacity(struct sched_domain *sd, int cpu)
7168 struct sched_domain *child = sd->child;
7169 struct sched_group *group, *sdg = sd->groups;
7170 unsigned long capacity, max_capacity, min_capacity;
7171 unsigned long interval;
7173 interval = msecs_to_jiffies(sd->balance_interval);
7174 interval = clamp(interval, 1UL, max_load_balance_interval);
7175 sdg->sgc->next_update = jiffies + interval;
7178 update_cpu_capacity(sd, cpu);
7184 min_capacity = ULONG_MAX;
7186 if (child->flags & SD_OVERLAP) {
7188 * SD_OVERLAP domains cannot assume that child groups
7189 * span the current group.
7192 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7193 struct sched_group_capacity *sgc;
7194 struct rq *rq = cpu_rq(cpu);
7197 * build_sched_domains() -> init_sched_groups_capacity()
7198 * gets here before we've attached the domains to the
7201 * Use capacity_of(), which is set irrespective of domains
7202 * in update_cpu_capacity().
7204 * This avoids capacity from being 0 and
7205 * causing divide-by-zero issues on boot.
7207 if (unlikely(!rq->sd)) {
7208 capacity += capacity_of(cpu);
7210 sgc = rq->sd->groups->sgc;
7211 capacity += sgc->capacity;
7214 max_capacity = max(capacity, max_capacity);
7215 min_capacity = min(capacity, min_capacity);
7219 * !SD_OVERLAP domains can assume that child groups
7220 * span the current group.
7223 group = child->groups;
7225 struct sched_group_capacity *sgc = group->sgc;
7227 capacity += sgc->capacity;
7228 max_capacity = max(sgc->max_capacity, max_capacity);
7229 min_capacity = min(sgc->min_capacity, min_capacity);
7230 group = group->next;
7231 } while (group != child->groups);
7234 sdg->sgc->capacity = capacity;
7235 sdg->sgc->max_capacity = max_capacity;
7236 sdg->sgc->min_capacity = min_capacity;
7240 * Check whether the capacity of the rq has been noticeably reduced by side
7241 * activity. The imbalance_pct is used for the threshold.
7242 * Return true is the capacity is reduced
7245 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7247 return ((rq->cpu_capacity * sd->imbalance_pct) <
7248 (rq->cpu_capacity_orig * 100));
7252 * Group imbalance indicates (and tries to solve) the problem where balancing
7253 * groups is inadequate due to tsk_cpus_allowed() constraints.
7255 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7256 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7259 * { 0 1 2 3 } { 4 5 6 7 }
7262 * If we were to balance group-wise we'd place two tasks in the first group and
7263 * two tasks in the second group. Clearly this is undesired as it will overload
7264 * cpu 3 and leave one of the cpus in the second group unused.
7266 * The current solution to this issue is detecting the skew in the first group
7267 * by noticing the lower domain failed to reach balance and had difficulty
7268 * moving tasks due to affinity constraints.
7270 * When this is so detected; this group becomes a candidate for busiest; see
7271 * update_sd_pick_busiest(). And calculate_imbalance() and
7272 * find_busiest_group() avoid some of the usual balance conditions to allow it
7273 * to create an effective group imbalance.
7275 * This is a somewhat tricky proposition since the next run might not find the
7276 * group imbalance and decide the groups need to be balanced again. A most
7277 * subtle and fragile situation.
7280 static inline int sg_imbalanced(struct sched_group *group)
7282 return group->sgc->imbalance;
7286 * group_has_capacity returns true if the group has spare capacity that could
7287 * be used by some tasks.
7288 * We consider that a group has spare capacity if the * number of task is
7289 * smaller than the number of CPUs or if the utilization is lower than the
7290 * available capacity for CFS tasks.
7291 * For the latter, we use a threshold to stabilize the state, to take into
7292 * account the variance of the tasks' load and to return true if the available
7293 * capacity in meaningful for the load balancer.
7294 * As an example, an available capacity of 1% can appear but it doesn't make
7295 * any benefit for the load balance.
7298 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7300 if (sgs->sum_nr_running < sgs->group_weight)
7303 if ((sgs->group_capacity * 100) >
7304 (sgs->group_util * env->sd->imbalance_pct))
7311 * group_is_overloaded returns true if the group has more tasks than it can
7313 * group_is_overloaded is not equals to !group_has_capacity because a group
7314 * with the exact right number of tasks, has no more spare capacity but is not
7315 * overloaded so both group_has_capacity and group_is_overloaded return
7319 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7321 if (sgs->sum_nr_running <= sgs->group_weight)
7324 if ((sgs->group_capacity * 100) <
7325 (sgs->group_util * env->sd->imbalance_pct))
7333 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7334 * per-cpu capacity than sched_group ref.
7337 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7339 return sg->sgc->max_capacity + capacity_margin - SCHED_LOAD_SCALE <
7340 ref->sgc->max_capacity;
7344 group_type group_classify(struct sched_group *group,
7345 struct sg_lb_stats *sgs)
7347 if (sgs->group_no_capacity)
7348 return group_overloaded;
7350 if (sg_imbalanced(group))
7351 return group_imbalanced;
7353 if (sgs->group_misfit_task)
7354 return group_misfit_task;
7360 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7361 * @env: The load balancing environment.
7362 * @group: sched_group whose statistics are to be updated.
7363 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7364 * @local_group: Does group contain this_cpu.
7365 * @sgs: variable to hold the statistics for this group.
7366 * @overload: Indicate more than one runnable task for any CPU.
7367 * @overutilized: Indicate overutilization for any CPU.
7369 static inline void update_sg_lb_stats(struct lb_env *env,
7370 struct sched_group *group, int load_idx,
7371 int local_group, struct sg_lb_stats *sgs,
7372 bool *overload, bool *overutilized)
7377 memset(sgs, 0, sizeof(*sgs));
7379 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7380 struct rq *rq = cpu_rq(i);
7382 /* Bias balancing toward cpus of our domain */
7384 load = target_load(i, load_idx);
7386 load = source_load(i, load_idx);
7388 sgs->group_load += load;
7389 sgs->group_util += cpu_util(i);
7390 sgs->sum_nr_running += rq->cfs.h_nr_running;
7392 nr_running = rq->nr_running;
7396 #ifdef CONFIG_NUMA_BALANCING
7397 sgs->nr_numa_running += rq->nr_numa_running;
7398 sgs->nr_preferred_running += rq->nr_preferred_running;
7400 sgs->sum_weighted_load += weighted_cpuload(i);
7402 * No need to call idle_cpu() if nr_running is not 0
7404 if (!nr_running && idle_cpu(i))
7407 if (cpu_overutilized(i)) {
7408 *overutilized = true;
7409 if (!sgs->group_misfit_task && rq->misfit_task)
7410 sgs->group_misfit_task = capacity_of(i);
7414 /* Adjust by relative CPU capacity of the group */
7415 sgs->group_capacity = group->sgc->capacity;
7416 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7418 if (sgs->sum_nr_running)
7419 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7421 sgs->group_weight = group->group_weight;
7423 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7424 sgs->group_type = group_classify(group, sgs);
7428 * update_sd_pick_busiest - return 1 on busiest group
7429 * @env: The load balancing environment.
7430 * @sds: sched_domain statistics
7431 * @sg: sched_group candidate to be checked for being the busiest
7432 * @sgs: sched_group statistics
7434 * Determine if @sg is a busier group than the previously selected
7437 * Return: %true if @sg is a busier group than the previously selected
7438 * busiest group. %false otherwise.
7440 static bool update_sd_pick_busiest(struct lb_env *env,
7441 struct sd_lb_stats *sds,
7442 struct sched_group *sg,
7443 struct sg_lb_stats *sgs)
7445 struct sg_lb_stats *busiest = &sds->busiest_stat;
7447 if (sgs->group_type > busiest->group_type)
7450 if (sgs->group_type < busiest->group_type)
7454 * Candidate sg doesn't face any serious load-balance problems
7455 * so don't pick it if the local sg is already filled up.
7457 if (sgs->group_type == group_other &&
7458 !group_has_capacity(env, &sds->local_stat))
7461 if (sgs->avg_load <= busiest->avg_load)
7464 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7468 * Candidate sg has no more than one task per CPU and
7469 * has higher per-CPU capacity. Migrating tasks to less
7470 * capable CPUs may harm throughput. Maximize throughput,
7471 * power/energy consequences are not considered.
7473 if (sgs->sum_nr_running <= sgs->group_weight &&
7474 group_smaller_cpu_capacity(sds->local, sg))
7478 /* This is the busiest node in its class. */
7479 if (!(env->sd->flags & SD_ASYM_PACKING))
7483 * ASYM_PACKING needs to move all the work to the lowest
7484 * numbered CPUs in the group, therefore mark all groups
7485 * higher than ourself as busy.
7487 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
7491 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
7498 #ifdef CONFIG_NUMA_BALANCING
7499 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7501 if (sgs->sum_nr_running > sgs->nr_numa_running)
7503 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7508 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7510 if (rq->nr_running > rq->nr_numa_running)
7512 if (rq->nr_running > rq->nr_preferred_running)
7517 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7522 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7526 #endif /* CONFIG_NUMA_BALANCING */
7529 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7530 * @env: The load balancing environment.
7531 * @sds: variable to hold the statistics for this sched_domain.
7533 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7535 struct sched_domain *child = env->sd->child;
7536 struct sched_group *sg = env->sd->groups;
7537 struct sg_lb_stats tmp_sgs;
7538 int load_idx, prefer_sibling = 0;
7539 bool overload = false, overutilized = false;
7541 if (child && child->flags & SD_PREFER_SIBLING)
7544 load_idx = get_sd_load_idx(env->sd, env->idle);
7547 struct sg_lb_stats *sgs = &tmp_sgs;
7550 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7553 sgs = &sds->local_stat;
7555 if (env->idle != CPU_NEWLY_IDLE ||
7556 time_after_eq(jiffies, sg->sgc->next_update))
7557 update_group_capacity(env->sd, env->dst_cpu);
7560 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7561 &overload, &overutilized);
7567 * In case the child domain prefers tasks go to siblings
7568 * first, lower the sg capacity so that we'll try
7569 * and move all the excess tasks away. We lower the capacity
7570 * of a group only if the local group has the capacity to fit
7571 * these excess tasks. The extra check prevents the case where
7572 * you always pull from the heaviest group when it is already
7573 * under-utilized (possible with a large weight task outweighs
7574 * the tasks on the system).
7576 if (prefer_sibling && sds->local &&
7577 group_has_capacity(env, &sds->local_stat) &&
7578 (sgs->sum_nr_running > 1)) {
7579 sgs->group_no_capacity = 1;
7580 sgs->group_type = group_classify(sg, sgs);
7584 * Ignore task groups with misfit tasks if local group has no
7585 * capacity or if per-cpu capacity isn't higher.
7587 if (sgs->group_type == group_misfit_task &&
7588 (!group_has_capacity(env, &sds->local_stat) ||
7589 !group_smaller_cpu_capacity(sg, sds->local)))
7590 sgs->group_type = group_other;
7592 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7594 sds->busiest_stat = *sgs;
7598 /* Now, start updating sd_lb_stats */
7599 sds->total_load += sgs->group_load;
7600 sds->total_capacity += sgs->group_capacity;
7603 } while (sg != env->sd->groups);
7605 if (env->sd->flags & SD_NUMA)
7606 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7608 env->src_grp_nr_running = sds->busiest_stat.sum_nr_running;
7610 if (!env->sd->parent) {
7611 /* update overload indicator if we are at root domain */
7612 if (env->dst_rq->rd->overload != overload)
7613 env->dst_rq->rd->overload = overload;
7615 /* Update over-utilization (tipping point, U >= 0) indicator */
7616 if (env->dst_rq->rd->overutilized != overutilized) {
7617 env->dst_rq->rd->overutilized = overutilized;
7618 trace_sched_overutilized(overutilized);
7621 if (!env->dst_rq->rd->overutilized && overutilized) {
7622 env->dst_rq->rd->overutilized = true;
7623 trace_sched_overutilized(true);
7630 * check_asym_packing - Check to see if the group is packed into the
7633 * This is primarily intended to used at the sibling level. Some
7634 * cores like POWER7 prefer to use lower numbered SMT threads. In the
7635 * case of POWER7, it can move to lower SMT modes only when higher
7636 * threads are idle. When in lower SMT modes, the threads will
7637 * perform better since they share less core resources. Hence when we
7638 * have idle threads, we want them to be the higher ones.
7640 * This packing function is run on idle threads. It checks to see if
7641 * the busiest CPU in this domain (core in the P7 case) has a higher
7642 * CPU number than the packing function is being run on. Here we are
7643 * assuming lower CPU number will be equivalent to lower a SMT thread
7646 * Return: 1 when packing is required and a task should be moved to
7647 * this CPU. The amount of the imbalance is returned in *imbalance.
7649 * @env: The load balancing environment.
7650 * @sds: Statistics of the sched_domain which is to be packed
7652 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7656 if (!(env->sd->flags & SD_ASYM_PACKING))
7662 busiest_cpu = group_first_cpu(sds->busiest);
7663 if (env->dst_cpu > busiest_cpu)
7666 env->imbalance = DIV_ROUND_CLOSEST(
7667 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7668 SCHED_CAPACITY_SCALE);
7674 * fix_small_imbalance - Calculate the minor imbalance that exists
7675 * amongst the groups of a sched_domain, during
7677 * @env: The load balancing environment.
7678 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7681 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7683 unsigned long tmp, capa_now = 0, capa_move = 0;
7684 unsigned int imbn = 2;
7685 unsigned long scaled_busy_load_per_task;
7686 struct sg_lb_stats *local, *busiest;
7688 local = &sds->local_stat;
7689 busiest = &sds->busiest_stat;
7691 if (!local->sum_nr_running)
7692 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7693 else if (busiest->load_per_task > local->load_per_task)
7696 scaled_busy_load_per_task =
7697 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7698 busiest->group_capacity;
7700 if (busiest->avg_load + scaled_busy_load_per_task >=
7701 local->avg_load + (scaled_busy_load_per_task * imbn)) {
7702 env->imbalance = busiest->load_per_task;
7707 * OK, we don't have enough imbalance to justify moving tasks,
7708 * however we may be able to increase total CPU capacity used by
7712 capa_now += busiest->group_capacity *
7713 min(busiest->load_per_task, busiest->avg_load);
7714 capa_now += local->group_capacity *
7715 min(local->load_per_task, local->avg_load);
7716 capa_now /= SCHED_CAPACITY_SCALE;
7718 /* Amount of load we'd subtract */
7719 if (busiest->avg_load > scaled_busy_load_per_task) {
7720 capa_move += busiest->group_capacity *
7721 min(busiest->load_per_task,
7722 busiest->avg_load - scaled_busy_load_per_task);
7725 /* Amount of load we'd add */
7726 if (busiest->avg_load * busiest->group_capacity <
7727 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7728 tmp = (busiest->avg_load * busiest->group_capacity) /
7729 local->group_capacity;
7731 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7732 local->group_capacity;
7734 capa_move += local->group_capacity *
7735 min(local->load_per_task, local->avg_load + tmp);
7736 capa_move /= SCHED_CAPACITY_SCALE;
7738 /* Move if we gain throughput */
7739 if (capa_move > capa_now)
7740 env->imbalance = busiest->load_per_task;
7744 * calculate_imbalance - Calculate the amount of imbalance present within the
7745 * groups of a given sched_domain during load balance.
7746 * @env: load balance environment
7747 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7749 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7751 unsigned long max_pull, load_above_capacity = ~0UL;
7752 struct sg_lb_stats *local, *busiest;
7754 local = &sds->local_stat;
7755 busiest = &sds->busiest_stat;
7757 if (busiest->group_type == group_imbalanced) {
7759 * In the group_imb case we cannot rely on group-wide averages
7760 * to ensure cpu-load equilibrium, look at wider averages. XXX
7762 busiest->load_per_task =
7763 min(busiest->load_per_task, sds->avg_load);
7767 * In the presence of smp nice balancing, certain scenarios can have
7768 * max load less than avg load(as we skip the groups at or below
7769 * its cpu_capacity, while calculating max_load..)
7771 if (busiest->avg_load <= sds->avg_load ||
7772 local->avg_load >= sds->avg_load) {
7773 /* Misfitting tasks should be migrated in any case */
7774 if (busiest->group_type == group_misfit_task) {
7775 env->imbalance = busiest->group_misfit_task;
7780 * Busiest group is overloaded, local is not, use the spare
7781 * cycles to maximize throughput
7783 if (busiest->group_type == group_overloaded &&
7784 local->group_type <= group_misfit_task) {
7785 env->imbalance = busiest->load_per_task;
7790 return fix_small_imbalance(env, sds);
7794 * If there aren't any idle cpus, avoid creating some.
7796 if (busiest->group_type == group_overloaded &&
7797 local->group_type == group_overloaded) {
7798 load_above_capacity = busiest->sum_nr_running *
7800 if (load_above_capacity > busiest->group_capacity)
7801 load_above_capacity -= busiest->group_capacity;
7803 load_above_capacity = ~0UL;
7807 * We're trying to get all the cpus to the average_load, so we don't
7808 * want to push ourselves above the average load, nor do we wish to
7809 * reduce the max loaded cpu below the average load. At the same time,
7810 * we also don't want to reduce the group load below the group capacity
7811 * (so that we can implement power-savings policies etc). Thus we look
7812 * for the minimum possible imbalance.
7814 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7816 /* How much load to actually move to equalise the imbalance */
7817 env->imbalance = min(
7818 max_pull * busiest->group_capacity,
7819 (sds->avg_load - local->avg_load) * local->group_capacity
7820 ) / SCHED_CAPACITY_SCALE;
7822 /* Boost imbalance to allow misfit task to be balanced. */
7823 if (busiest->group_type == group_misfit_task)
7824 env->imbalance = max_t(long, env->imbalance,
7825 busiest->group_misfit_task);
7828 * if *imbalance is less than the average load per runnable task
7829 * there is no guarantee that any tasks will be moved so we'll have
7830 * a think about bumping its value to force at least one task to be
7833 if (env->imbalance < busiest->load_per_task)
7834 return fix_small_imbalance(env, sds);
7837 /******* find_busiest_group() helpers end here *********************/
7840 * find_busiest_group - Returns the busiest group within the sched_domain
7841 * if there is an imbalance. If there isn't an imbalance, and
7842 * the user has opted for power-savings, it returns a group whose
7843 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
7844 * such a group exists.
7846 * Also calculates the amount of weighted load which should be moved
7847 * to restore balance.
7849 * @env: The load balancing environment.
7851 * Return: - The busiest group if imbalance exists.
7852 * - If no imbalance and user has opted for power-savings balance,
7853 * return the least loaded group whose CPUs can be
7854 * put to idle by rebalancing its tasks onto our group.
7856 static struct sched_group *find_busiest_group(struct lb_env *env)
7858 struct sg_lb_stats *local, *busiest;
7859 struct sd_lb_stats sds;
7861 init_sd_lb_stats(&sds);
7864 * Compute the various statistics relavent for load balancing at
7867 update_sd_lb_stats(env, &sds);
7869 if (energy_aware() && !env->dst_rq->rd->overutilized)
7872 local = &sds.local_stat;
7873 busiest = &sds.busiest_stat;
7875 /* ASYM feature bypasses nice load balance check */
7876 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
7877 check_asym_packing(env, &sds))
7880 /* There is no busy sibling group to pull tasks from */
7881 if (!sds.busiest || busiest->sum_nr_running == 0)
7884 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7885 / sds.total_capacity;
7888 * If the busiest group is imbalanced the below checks don't
7889 * work because they assume all things are equal, which typically
7890 * isn't true due to cpus_allowed constraints and the like.
7892 if (busiest->group_type == group_imbalanced)
7895 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7896 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7897 busiest->group_no_capacity)
7900 /* Misfitting tasks should be dealt with regardless of the avg load */
7901 if (busiest->group_type == group_misfit_task) {
7906 * If the local group is busier than the selected busiest group
7907 * don't try and pull any tasks.
7909 if (local->avg_load >= busiest->avg_load)
7913 * Don't pull any tasks if this group is already above the domain
7916 if (local->avg_load >= sds.avg_load)
7919 if (env->idle == CPU_IDLE) {
7921 * This cpu is idle. If the busiest group is not overloaded
7922 * and there is no imbalance between this and busiest group
7923 * wrt idle cpus, it is balanced. The imbalance becomes
7924 * significant if the diff is greater than 1 otherwise we
7925 * might end up to just move the imbalance on another group
7927 if ((busiest->group_type != group_overloaded) &&
7928 (local->idle_cpus <= (busiest->idle_cpus + 1)) &&
7929 !group_smaller_cpu_capacity(sds.busiest, sds.local))
7933 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7934 * imbalance_pct to be conservative.
7936 if (100 * busiest->avg_load <=
7937 env->sd->imbalance_pct * local->avg_load)
7942 env->busiest_group_type = busiest->group_type;
7943 /* Looks like there is an imbalance. Compute it */
7944 calculate_imbalance(env, &sds);
7953 * find_busiest_queue - find the busiest runqueue among the cpus in group.
7955 static struct rq *find_busiest_queue(struct lb_env *env,
7956 struct sched_group *group)
7958 struct rq *busiest = NULL, *rq;
7959 unsigned long busiest_load = 0, busiest_capacity = 1;
7962 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7963 unsigned long capacity, wl;
7967 rt = fbq_classify_rq(rq);
7970 * We classify groups/runqueues into three groups:
7971 * - regular: there are !numa tasks
7972 * - remote: there are numa tasks that run on the 'wrong' node
7973 * - all: there is no distinction
7975 * In order to avoid migrating ideally placed numa tasks,
7976 * ignore those when there's better options.
7978 * If we ignore the actual busiest queue to migrate another
7979 * task, the next balance pass can still reduce the busiest
7980 * queue by moving tasks around inside the node.
7982 * If we cannot move enough load due to this classification
7983 * the next pass will adjust the group classification and
7984 * allow migration of more tasks.
7986 * Both cases only affect the total convergence complexity.
7988 if (rt > env->fbq_type)
7991 capacity = capacity_of(i);
7993 wl = weighted_cpuload(i);
7996 * When comparing with imbalance, use weighted_cpuload()
7997 * which is not scaled with the cpu capacity.
8000 if (rq->nr_running == 1 && wl > env->imbalance &&
8001 !check_cpu_capacity(rq, env->sd) &&
8002 env->busiest_group_type != group_misfit_task)
8006 * For the load comparisons with the other cpu's, consider
8007 * the weighted_cpuload() scaled with the cpu capacity, so
8008 * that the load can be moved away from the cpu that is
8009 * potentially running at a lower capacity.
8011 * Thus we're looking for max(wl_i / capacity_i), crosswise
8012 * multiplication to rid ourselves of the division works out
8013 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8014 * our previous maximum.
8016 if (wl * busiest_capacity > busiest_load * capacity) {
8018 busiest_capacity = capacity;
8027 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8028 * so long as it is large enough.
8030 #define MAX_PINNED_INTERVAL 512
8032 /* Working cpumask for load_balance and load_balance_newidle. */
8033 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
8035 static int need_active_balance(struct lb_env *env)
8037 struct sched_domain *sd = env->sd;
8039 if (env->idle == CPU_NEWLY_IDLE) {
8042 * ASYM_PACKING needs to force migrate tasks from busy but
8043 * higher numbered CPUs in order to pack all tasks in the
8044 * lowest numbered CPUs.
8046 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
8051 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8052 * It's worth migrating the task if the src_cpu's capacity is reduced
8053 * because of other sched_class or IRQs if more capacity stays
8054 * available on dst_cpu.
8056 if ((env->idle != CPU_NOT_IDLE) &&
8057 (env->src_rq->cfs.h_nr_running == 1)) {
8058 if ((check_cpu_capacity(env->src_rq, sd)) &&
8059 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8063 if ((capacity_of(env->src_cpu) < capacity_of(env->dst_cpu)) &&
8064 env->src_rq->cfs.h_nr_running == 1 &&
8065 cpu_overutilized(env->src_cpu) &&
8066 !cpu_overutilized(env->dst_cpu)) {
8070 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8073 static int active_load_balance_cpu_stop(void *data);
8075 static int should_we_balance(struct lb_env *env)
8077 struct sched_group *sg = env->sd->groups;
8078 struct cpumask *sg_cpus, *sg_mask;
8079 int cpu, balance_cpu = -1;
8082 * In the newly idle case, we will allow all the cpu's
8083 * to do the newly idle load balance.
8085 if (env->idle == CPU_NEWLY_IDLE)
8088 sg_cpus = sched_group_cpus(sg);
8089 sg_mask = sched_group_mask(sg);
8090 /* Try to find first idle cpu */
8091 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
8092 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
8099 if (balance_cpu == -1)
8100 balance_cpu = group_balance_cpu(sg);
8103 * First idle cpu or the first cpu(busiest) in this sched group
8104 * is eligible for doing load balancing at this and above domains.
8106 return balance_cpu == env->dst_cpu;
8110 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8111 * tasks if there is an imbalance.
8113 static int load_balance(int this_cpu, struct rq *this_rq,
8114 struct sched_domain *sd, enum cpu_idle_type idle,
8115 int *continue_balancing)
8117 int ld_moved, cur_ld_moved, active_balance = 0;
8118 struct sched_domain *sd_parent = sd->parent;
8119 struct sched_group *group;
8121 unsigned long flags;
8122 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8124 struct lb_env env = {
8126 .dst_cpu = this_cpu,
8128 .dst_grpmask = sched_group_cpus(sd->groups),
8130 .loop_break = sched_nr_migrate_break,
8133 .tasks = LIST_HEAD_INIT(env.tasks),
8137 * For NEWLY_IDLE load_balancing, we don't need to consider
8138 * other cpus in our group
8140 if (idle == CPU_NEWLY_IDLE)
8141 env.dst_grpmask = NULL;
8143 cpumask_copy(cpus, cpu_active_mask);
8145 schedstat_inc(sd, lb_count[idle]);
8148 if (!should_we_balance(&env)) {
8149 *continue_balancing = 0;
8153 group = find_busiest_group(&env);
8155 schedstat_inc(sd, lb_nobusyg[idle]);
8159 busiest = find_busiest_queue(&env, group);
8161 schedstat_inc(sd, lb_nobusyq[idle]);
8165 BUG_ON(busiest == env.dst_rq);
8167 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
8169 env.src_cpu = busiest->cpu;
8170 env.src_rq = busiest;
8173 if (busiest->nr_running > 1) {
8175 * Attempt to move tasks. If find_busiest_group has found
8176 * an imbalance but busiest->nr_running <= 1, the group is
8177 * still unbalanced. ld_moved simply stays zero, so it is
8178 * correctly treated as an imbalance.
8180 env.flags |= LBF_ALL_PINNED;
8181 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8184 raw_spin_lock_irqsave(&busiest->lock, flags);
8187 * cur_ld_moved - load moved in current iteration
8188 * ld_moved - cumulative load moved across iterations
8190 cur_ld_moved = detach_tasks(&env);
8192 * We want to potentially lower env.src_cpu's OPP.
8195 update_capacity_of(env.src_cpu);
8198 * We've detached some tasks from busiest_rq. Every
8199 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8200 * unlock busiest->lock, and we are able to be sure
8201 * that nobody can manipulate the tasks in parallel.
8202 * See task_rq_lock() family for the details.
8205 raw_spin_unlock(&busiest->lock);
8209 ld_moved += cur_ld_moved;
8212 local_irq_restore(flags);
8214 if (env.flags & LBF_NEED_BREAK) {
8215 env.flags &= ~LBF_NEED_BREAK;
8220 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8221 * us and move them to an alternate dst_cpu in our sched_group
8222 * where they can run. The upper limit on how many times we
8223 * iterate on same src_cpu is dependent on number of cpus in our
8226 * This changes load balance semantics a bit on who can move
8227 * load to a given_cpu. In addition to the given_cpu itself
8228 * (or a ilb_cpu acting on its behalf where given_cpu is
8229 * nohz-idle), we now have balance_cpu in a position to move
8230 * load to given_cpu. In rare situations, this may cause
8231 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8232 * _independently_ and at _same_ time to move some load to
8233 * given_cpu) causing exceess load to be moved to given_cpu.
8234 * This however should not happen so much in practice and
8235 * moreover subsequent load balance cycles should correct the
8236 * excess load moved.
8238 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8240 /* Prevent to re-select dst_cpu via env's cpus */
8241 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8243 env.dst_rq = cpu_rq(env.new_dst_cpu);
8244 env.dst_cpu = env.new_dst_cpu;
8245 env.flags &= ~LBF_DST_PINNED;
8247 env.loop_break = sched_nr_migrate_break;
8250 * Go back to "more_balance" rather than "redo" since we
8251 * need to continue with same src_cpu.
8257 * We failed to reach balance because of affinity.
8260 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8262 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8263 *group_imbalance = 1;
8266 /* All tasks on this runqueue were pinned by CPU affinity */
8267 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8268 cpumask_clear_cpu(cpu_of(busiest), cpus);
8269 if (!cpumask_empty(cpus)) {
8271 env.loop_break = sched_nr_migrate_break;
8274 goto out_all_pinned;
8279 schedstat_inc(sd, lb_failed[idle]);
8281 * Increment the failure counter only on periodic balance.
8282 * We do not want newidle balance, which can be very
8283 * frequent, pollute the failure counter causing
8284 * excessive cache_hot migrations and active balances.
8286 if (idle != CPU_NEWLY_IDLE)
8287 if (env.src_grp_nr_running > 1)
8288 sd->nr_balance_failed++;
8290 if (need_active_balance(&env)) {
8291 raw_spin_lock_irqsave(&busiest->lock, flags);
8293 /* don't kick the active_load_balance_cpu_stop,
8294 * if the curr task on busiest cpu can't be
8297 if (!cpumask_test_cpu(this_cpu,
8298 tsk_cpus_allowed(busiest->curr))) {
8299 raw_spin_unlock_irqrestore(&busiest->lock,
8301 env.flags |= LBF_ALL_PINNED;
8302 goto out_one_pinned;
8306 * ->active_balance synchronizes accesses to
8307 * ->active_balance_work. Once set, it's cleared
8308 * only after active load balance is finished.
8310 if (!busiest->active_balance) {
8311 busiest->active_balance = 1;
8312 busiest->push_cpu = this_cpu;
8315 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8317 if (active_balance) {
8318 stop_one_cpu_nowait(cpu_of(busiest),
8319 active_load_balance_cpu_stop, busiest,
8320 &busiest->active_balance_work);
8324 * We've kicked active balancing, reset the failure
8327 sd->nr_balance_failed = sd->cache_nice_tries+1;
8330 sd->nr_balance_failed = 0;
8332 if (likely(!active_balance)) {
8333 /* We were unbalanced, so reset the balancing interval */
8334 sd->balance_interval = sd->min_interval;
8337 * If we've begun active balancing, start to back off. This
8338 * case may not be covered by the all_pinned logic if there
8339 * is only 1 task on the busy runqueue (because we don't call
8342 if (sd->balance_interval < sd->max_interval)
8343 sd->balance_interval *= 2;
8350 * We reach balance although we may have faced some affinity
8351 * constraints. Clear the imbalance flag if it was set.
8354 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8356 if (*group_imbalance)
8357 *group_imbalance = 0;
8362 * We reach balance because all tasks are pinned at this level so
8363 * we can't migrate them. Let the imbalance flag set so parent level
8364 * can try to migrate them.
8366 schedstat_inc(sd, lb_balanced[idle]);
8368 sd->nr_balance_failed = 0;
8371 /* tune up the balancing interval */
8372 if (((env.flags & LBF_ALL_PINNED) &&
8373 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8374 (sd->balance_interval < sd->max_interval))
8375 sd->balance_interval *= 2;
8382 static inline unsigned long
8383 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8385 unsigned long interval = sd->balance_interval;
8388 interval *= sd->busy_factor;
8390 /* scale ms to jiffies */
8391 interval = msecs_to_jiffies(interval);
8392 interval = clamp(interval, 1UL, max_load_balance_interval);
8398 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
8400 unsigned long interval, next;
8402 interval = get_sd_balance_interval(sd, cpu_busy);
8403 next = sd->last_balance + interval;
8405 if (time_after(*next_balance, next))
8406 *next_balance = next;
8410 * idle_balance is called by schedule() if this_cpu is about to become
8411 * idle. Attempts to pull tasks from other CPUs.
8413 static int idle_balance(struct rq *this_rq)
8415 unsigned long next_balance = jiffies + HZ;
8416 int this_cpu = this_rq->cpu;
8417 struct sched_domain *sd;
8418 int pulled_task = 0;
8420 long removed_util=0;
8422 idle_enter_fair(this_rq);
8425 * We must set idle_stamp _before_ calling idle_balance(), such that we
8426 * measure the duration of idle_balance() as idle time.
8428 this_rq->idle_stamp = rq_clock(this_rq);
8430 if (!energy_aware() &&
8431 (this_rq->avg_idle < sysctl_sched_migration_cost ||
8432 !this_rq->rd->overload)) {
8434 sd = rcu_dereference_check_sched_domain(this_rq->sd);
8436 update_next_balance(sd, 0, &next_balance);
8442 raw_spin_unlock(&this_rq->lock);
8445 * If removed_util_avg is !0 we most probably migrated some task away
8446 * from this_cpu. In this case we might be willing to trigger an OPP
8447 * update, but we want to do so if we don't find anybody else to pull
8448 * here (we will trigger an OPP update with the pulled task's enqueue
8451 * Record removed_util before calling update_blocked_averages, and use
8452 * it below (before returning) to see if an OPP update is required.
8454 removed_util = atomic_long_read(&(this_rq->cfs).removed_util_avg);
8455 update_blocked_averages(this_cpu);
8457 for_each_domain(this_cpu, sd) {
8458 int continue_balancing = 1;
8459 u64 t0, domain_cost;
8461 if (!(sd->flags & SD_LOAD_BALANCE))
8464 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
8465 update_next_balance(sd, 0, &next_balance);
8469 if (sd->flags & SD_BALANCE_NEWIDLE) {
8470 t0 = sched_clock_cpu(this_cpu);
8472 pulled_task = load_balance(this_cpu, this_rq,
8474 &continue_balancing);
8476 domain_cost = sched_clock_cpu(this_cpu) - t0;
8477 if (domain_cost > sd->max_newidle_lb_cost)
8478 sd->max_newidle_lb_cost = domain_cost;
8480 curr_cost += domain_cost;
8483 update_next_balance(sd, 0, &next_balance);
8486 * Stop searching for tasks to pull if there are
8487 * now runnable tasks on this rq.
8489 if (pulled_task || this_rq->nr_running > 0)
8494 raw_spin_lock(&this_rq->lock);
8496 if (curr_cost > this_rq->max_idle_balance_cost)
8497 this_rq->max_idle_balance_cost = curr_cost;
8500 * While browsing the domains, we released the rq lock, a task could
8501 * have been enqueued in the meantime. Since we're not going idle,
8502 * pretend we pulled a task.
8504 if (this_rq->cfs.h_nr_running && !pulled_task)
8508 /* Move the next balance forward */
8509 if (time_after(this_rq->next_balance, next_balance))
8510 this_rq->next_balance = next_balance;
8512 /* Is there a task of a high priority class? */
8513 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
8517 idle_exit_fair(this_rq);
8518 this_rq->idle_stamp = 0;
8519 } else if (removed_util) {
8521 * No task pulled and someone has been migrated away.
8522 * Good case to trigger an OPP update.
8524 update_capacity_of(this_cpu);
8531 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
8532 * running tasks off the busiest CPU onto idle CPUs. It requires at
8533 * least 1 task to be running on each physical CPU where possible, and
8534 * avoids physical / logical imbalances.
8536 static int active_load_balance_cpu_stop(void *data)
8538 struct rq *busiest_rq = data;
8539 int busiest_cpu = cpu_of(busiest_rq);
8540 int target_cpu = busiest_rq->push_cpu;
8541 struct rq *target_rq = cpu_rq(target_cpu);
8542 struct sched_domain *sd;
8543 struct task_struct *p = NULL;
8545 raw_spin_lock_irq(&busiest_rq->lock);
8547 /* make sure the requested cpu hasn't gone down in the meantime */
8548 if (unlikely(busiest_cpu != smp_processor_id() ||
8549 !busiest_rq->active_balance))
8552 /* Is there any task to move? */
8553 if (busiest_rq->nr_running <= 1)
8557 * This condition is "impossible", if it occurs
8558 * we need to fix it. Originally reported by
8559 * Bjorn Helgaas on a 128-cpu setup.
8561 BUG_ON(busiest_rq == target_rq);
8563 /* Search for an sd spanning us and the target CPU. */
8565 for_each_domain(target_cpu, sd) {
8566 if ((sd->flags & SD_LOAD_BALANCE) &&
8567 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8572 struct lb_env env = {
8574 .dst_cpu = target_cpu,
8575 .dst_rq = target_rq,
8576 .src_cpu = busiest_rq->cpu,
8577 .src_rq = busiest_rq,
8581 schedstat_inc(sd, alb_count);
8583 p = detach_one_task(&env);
8585 schedstat_inc(sd, alb_pushed);
8587 * We want to potentially lower env.src_cpu's OPP.
8589 update_capacity_of(env.src_cpu);
8592 schedstat_inc(sd, alb_failed);
8596 busiest_rq->active_balance = 0;
8597 raw_spin_unlock(&busiest_rq->lock);
8600 attach_one_task(target_rq, p);
8607 static inline int on_null_domain(struct rq *rq)
8609 return unlikely(!rcu_dereference_sched(rq->sd));
8612 #ifdef CONFIG_NO_HZ_COMMON
8614 * idle load balancing details
8615 * - When one of the busy CPUs notice that there may be an idle rebalancing
8616 * needed, they will kick the idle load balancer, which then does idle
8617 * load balancing for all the idle CPUs.
8620 cpumask_var_t idle_cpus_mask;
8622 unsigned long next_balance; /* in jiffy units */
8623 } nohz ____cacheline_aligned;
8625 static inline int find_new_ilb(void)
8627 int ilb = cpumask_first(nohz.idle_cpus_mask);
8629 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8636 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8637 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8638 * CPU (if there is one).
8640 static void nohz_balancer_kick(void)
8644 nohz.next_balance++;
8646 ilb_cpu = find_new_ilb();
8648 if (ilb_cpu >= nr_cpu_ids)
8651 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
8654 * Use smp_send_reschedule() instead of resched_cpu().
8655 * This way we generate a sched IPI on the target cpu which
8656 * is idle. And the softirq performing nohz idle load balance
8657 * will be run before returning from the IPI.
8659 smp_send_reschedule(ilb_cpu);
8663 static inline void nohz_balance_exit_idle(int cpu)
8665 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
8667 * Completely isolated CPUs don't ever set, so we must test.
8669 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
8670 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
8671 atomic_dec(&nohz.nr_cpus);
8673 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8677 static inline void set_cpu_sd_state_busy(void)
8679 struct sched_domain *sd;
8680 int cpu = smp_processor_id();
8683 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8685 if (!sd || !sd->nohz_idle)
8689 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
8694 void set_cpu_sd_state_idle(void)
8696 struct sched_domain *sd;
8697 int cpu = smp_processor_id();
8700 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8702 if (!sd || sd->nohz_idle)
8706 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
8712 * This routine will record that the cpu is going idle with tick stopped.
8713 * This info will be used in performing idle load balancing in the future.
8715 void nohz_balance_enter_idle(int cpu)
8718 * If this cpu is going down, then nothing needs to be done.
8720 if (!cpu_active(cpu))
8723 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8727 * If we're a completely isolated CPU, we don't play.
8729 if (on_null_domain(cpu_rq(cpu)))
8732 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8733 atomic_inc(&nohz.nr_cpus);
8734 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8737 static int sched_ilb_notifier(struct notifier_block *nfb,
8738 unsigned long action, void *hcpu)
8740 switch (action & ~CPU_TASKS_FROZEN) {
8742 nohz_balance_exit_idle(smp_processor_id());
8750 static DEFINE_SPINLOCK(balancing);
8753 * Scale the max load_balance interval with the number of CPUs in the system.
8754 * This trades load-balance latency on larger machines for less cross talk.
8756 void update_max_interval(void)
8758 max_load_balance_interval = HZ*num_online_cpus()/10;
8762 * It checks each scheduling domain to see if it is due to be balanced,
8763 * and initiates a balancing operation if so.
8765 * Balancing parameters are set up in init_sched_domains.
8767 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8769 int continue_balancing = 1;
8771 unsigned long interval;
8772 struct sched_domain *sd;
8773 /* Earliest time when we have to do rebalance again */
8774 unsigned long next_balance = jiffies + 60*HZ;
8775 int update_next_balance = 0;
8776 int need_serialize, need_decay = 0;
8779 update_blocked_averages(cpu);
8782 for_each_domain(cpu, sd) {
8784 * Decay the newidle max times here because this is a regular
8785 * visit to all the domains. Decay ~1% per second.
8787 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8788 sd->max_newidle_lb_cost =
8789 (sd->max_newidle_lb_cost * 253) / 256;
8790 sd->next_decay_max_lb_cost = jiffies + HZ;
8793 max_cost += sd->max_newidle_lb_cost;
8795 if (!(sd->flags & SD_LOAD_BALANCE))
8799 * Stop the load balance at this level. There is another
8800 * CPU in our sched group which is doing load balancing more
8803 if (!continue_balancing) {
8809 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8811 need_serialize = sd->flags & SD_SERIALIZE;
8812 if (need_serialize) {
8813 if (!spin_trylock(&balancing))
8817 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8818 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8820 * The LBF_DST_PINNED logic could have changed
8821 * env->dst_cpu, so we can't know our idle
8822 * state even if we migrated tasks. Update it.
8824 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8826 sd->last_balance = jiffies;
8827 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8830 spin_unlock(&balancing);
8832 if (time_after(next_balance, sd->last_balance + interval)) {
8833 next_balance = sd->last_balance + interval;
8834 update_next_balance = 1;
8839 * Ensure the rq-wide value also decays but keep it at a
8840 * reasonable floor to avoid funnies with rq->avg_idle.
8842 rq->max_idle_balance_cost =
8843 max((u64)sysctl_sched_migration_cost, max_cost);
8848 * next_balance will be updated only when there is a need.
8849 * When the cpu is attached to null domain for ex, it will not be
8852 if (likely(update_next_balance)) {
8853 rq->next_balance = next_balance;
8855 #ifdef CONFIG_NO_HZ_COMMON
8857 * If this CPU has been elected to perform the nohz idle
8858 * balance. Other idle CPUs have already rebalanced with
8859 * nohz_idle_balance() and nohz.next_balance has been
8860 * updated accordingly. This CPU is now running the idle load
8861 * balance for itself and we need to update the
8862 * nohz.next_balance accordingly.
8864 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8865 nohz.next_balance = rq->next_balance;
8870 #ifdef CONFIG_NO_HZ_COMMON
8872 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8873 * rebalancing for all the cpus for whom scheduler ticks are stopped.
8875 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8877 int this_cpu = this_rq->cpu;
8880 /* Earliest time when we have to do rebalance again */
8881 unsigned long next_balance = jiffies + 60*HZ;
8882 int update_next_balance = 0;
8884 if (idle != CPU_IDLE ||
8885 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8888 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8889 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8893 * If this cpu gets work to do, stop the load balancing
8894 * work being done for other cpus. Next load
8895 * balancing owner will pick it up.
8900 rq = cpu_rq(balance_cpu);
8903 * If time for next balance is due,
8906 if (time_after_eq(jiffies, rq->next_balance)) {
8907 raw_spin_lock_irq(&rq->lock);
8908 update_rq_clock(rq);
8909 update_idle_cpu_load(rq);
8910 raw_spin_unlock_irq(&rq->lock);
8911 rebalance_domains(rq, CPU_IDLE);
8914 if (time_after(next_balance, rq->next_balance)) {
8915 next_balance = rq->next_balance;
8916 update_next_balance = 1;
8921 * next_balance will be updated only when there is a need.
8922 * When the CPU is attached to null domain for ex, it will not be
8925 if (likely(update_next_balance))
8926 nohz.next_balance = next_balance;
8928 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8932 * Current heuristic for kicking the idle load balancer in the presence
8933 * of an idle cpu in the system.
8934 * - This rq has more than one task.
8935 * - This rq has at least one CFS task and the capacity of the CPU is
8936 * significantly reduced because of RT tasks or IRQs.
8937 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
8938 * multiple busy cpu.
8939 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8940 * domain span are idle.
8942 static inline bool nohz_kick_needed(struct rq *rq)
8944 unsigned long now = jiffies;
8945 struct sched_domain *sd;
8946 struct sched_group_capacity *sgc;
8947 int nr_busy, cpu = rq->cpu;
8950 if (unlikely(rq->idle_balance))
8954 * We may be recently in ticked or tickless idle mode. At the first
8955 * busy tick after returning from idle, we will update the busy stats.
8957 set_cpu_sd_state_busy();
8958 nohz_balance_exit_idle(cpu);
8961 * None are in tickless mode and hence no need for NOHZ idle load
8964 if (likely(!atomic_read(&nohz.nr_cpus)))
8967 if (time_before(now, nohz.next_balance))
8970 if (rq->nr_running >= 2 &&
8971 (!energy_aware() || cpu_overutilized(cpu)))
8975 sd = rcu_dereference(per_cpu(sd_busy, cpu));
8976 if (sd && !energy_aware()) {
8977 sgc = sd->groups->sgc;
8978 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8987 sd = rcu_dereference(rq->sd);
8989 if ((rq->cfs.h_nr_running >= 1) &&
8990 check_cpu_capacity(rq, sd)) {
8996 sd = rcu_dereference(per_cpu(sd_asym, cpu));
8997 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8998 sched_domain_span(sd)) < cpu)) {
9008 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
9012 * run_rebalance_domains is triggered when needed from the scheduler tick.
9013 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9015 static void run_rebalance_domains(struct softirq_action *h)
9017 struct rq *this_rq = this_rq();
9018 enum cpu_idle_type idle = this_rq->idle_balance ?
9019 CPU_IDLE : CPU_NOT_IDLE;
9022 * If this cpu has a pending nohz_balance_kick, then do the
9023 * balancing on behalf of the other idle cpus whose ticks are
9024 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9025 * give the idle cpus a chance to load balance. Else we may
9026 * load balance only within the local sched_domain hierarchy
9027 * and abort nohz_idle_balance altogether if we pull some load.
9029 nohz_idle_balance(this_rq, idle);
9030 rebalance_domains(this_rq, idle);
9034 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9036 void trigger_load_balance(struct rq *rq)
9038 /* Don't need to rebalance while attached to NULL domain */
9039 if (unlikely(on_null_domain(rq)))
9042 if (time_after_eq(jiffies, rq->next_balance))
9043 raise_softirq(SCHED_SOFTIRQ);
9044 #ifdef CONFIG_NO_HZ_COMMON
9045 if (nohz_kick_needed(rq))
9046 nohz_balancer_kick();
9050 static void rq_online_fair(struct rq *rq)
9054 update_runtime_enabled(rq);
9057 static void rq_offline_fair(struct rq *rq)
9061 /* Ensure any throttled groups are reachable by pick_next_task */
9062 unthrottle_offline_cfs_rqs(rq);
9065 #endif /* CONFIG_SMP */
9068 * scheduler tick hitting a task of our scheduling class:
9070 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9072 struct cfs_rq *cfs_rq;
9073 struct sched_entity *se = &curr->se;
9075 for_each_sched_entity(se) {
9076 cfs_rq = cfs_rq_of(se);
9077 entity_tick(cfs_rq, se, queued);
9080 if (static_branch_unlikely(&sched_numa_balancing))
9081 task_tick_numa(rq, curr);
9084 if (!rq->rd->overutilized && cpu_overutilized(task_cpu(curr))) {
9085 rq->rd->overutilized = true;
9086 trace_sched_overutilized(true);
9089 rq->misfit_task = !task_fits_max(curr, rq->cpu);
9095 * called on fork with the child task as argument from the parent's context
9096 * - child not yet on the tasklist
9097 * - preemption disabled
9099 static void task_fork_fair(struct task_struct *p)
9101 struct cfs_rq *cfs_rq;
9102 struct sched_entity *se = &p->se, *curr;
9103 int this_cpu = smp_processor_id();
9104 struct rq *rq = this_rq();
9105 unsigned long flags;
9107 raw_spin_lock_irqsave(&rq->lock, flags);
9109 update_rq_clock(rq);
9111 cfs_rq = task_cfs_rq(current);
9112 curr = cfs_rq->curr;
9115 * Not only the cpu but also the task_group of the parent might have
9116 * been changed after parent->se.parent,cfs_rq were copied to
9117 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
9118 * of child point to valid ones.
9121 __set_task_cpu(p, this_cpu);
9124 update_curr(cfs_rq);
9127 se->vruntime = curr->vruntime;
9128 place_entity(cfs_rq, se, 1);
9130 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9132 * Upon rescheduling, sched_class::put_prev_task() will place
9133 * 'current' within the tree based on its new key value.
9135 swap(curr->vruntime, se->vruntime);
9139 se->vruntime -= cfs_rq->min_vruntime;
9141 raw_spin_unlock_irqrestore(&rq->lock, flags);
9145 * Priority of the task has changed. Check to see if we preempt
9149 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9151 if (!task_on_rq_queued(p))
9155 * Reschedule if we are currently running on this runqueue and
9156 * our priority decreased, or if we are not currently running on
9157 * this runqueue and our priority is higher than the current's
9159 if (rq->curr == p) {
9160 if (p->prio > oldprio)
9163 check_preempt_curr(rq, p, 0);
9166 static inline bool vruntime_normalized(struct task_struct *p)
9168 struct sched_entity *se = &p->se;
9171 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9172 * the dequeue_entity(.flags=0) will already have normalized the
9179 * When !on_rq, vruntime of the task has usually NOT been normalized.
9180 * But there are some cases where it has already been normalized:
9182 * - A forked child which is waiting for being woken up by
9183 * wake_up_new_task().
9184 * - A task which has been woken up by try_to_wake_up() and
9185 * waiting for actually being woken up by sched_ttwu_pending().
9187 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9193 static void detach_task_cfs_rq(struct task_struct *p)
9195 struct sched_entity *se = &p->se;
9196 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9198 if (!vruntime_normalized(p)) {
9200 * Fix up our vruntime so that the current sleep doesn't
9201 * cause 'unlimited' sleep bonus.
9203 place_entity(cfs_rq, se, 0);
9204 se->vruntime -= cfs_rq->min_vruntime;
9207 /* Catch up with the cfs_rq and remove our load when we leave */
9208 detach_entity_load_avg(cfs_rq, se);
9211 static void attach_task_cfs_rq(struct task_struct *p)
9213 struct sched_entity *se = &p->se;
9214 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9216 #ifdef CONFIG_FAIR_GROUP_SCHED
9218 * Since the real-depth could have been changed (only FAIR
9219 * class maintain depth value), reset depth properly.
9221 se->depth = se->parent ? se->parent->depth + 1 : 0;
9224 /* Synchronize task with its cfs_rq */
9225 attach_entity_load_avg(cfs_rq, se);
9227 if (!vruntime_normalized(p))
9228 se->vruntime += cfs_rq->min_vruntime;
9231 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9233 detach_task_cfs_rq(p);
9236 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9238 attach_task_cfs_rq(p);
9240 if (task_on_rq_queued(p)) {
9242 * We were most likely switched from sched_rt, so
9243 * kick off the schedule if running, otherwise just see
9244 * if we can still preempt the current task.
9249 check_preempt_curr(rq, p, 0);
9253 /* Account for a task changing its policy or group.
9255 * This routine is mostly called to set cfs_rq->curr field when a task
9256 * migrates between groups/classes.
9258 static void set_curr_task_fair(struct rq *rq)
9260 struct sched_entity *se = &rq->curr->se;
9262 for_each_sched_entity(se) {
9263 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9265 set_next_entity(cfs_rq, se);
9266 /* ensure bandwidth has been allocated on our new cfs_rq */
9267 account_cfs_rq_runtime(cfs_rq, 0);
9271 void init_cfs_rq(struct cfs_rq *cfs_rq)
9273 cfs_rq->tasks_timeline = RB_ROOT;
9274 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9275 #ifndef CONFIG_64BIT
9276 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9279 atomic_long_set(&cfs_rq->removed_load_avg, 0);
9280 atomic_long_set(&cfs_rq->removed_util_avg, 0);
9284 #ifdef CONFIG_FAIR_GROUP_SCHED
9285 static void task_move_group_fair(struct task_struct *p)
9287 detach_task_cfs_rq(p);
9288 set_task_rq(p, task_cpu(p));
9291 /* Tell se's cfs_rq has been changed -- migrated */
9292 p->se.avg.last_update_time = 0;
9294 attach_task_cfs_rq(p);
9297 void free_fair_sched_group(struct task_group *tg)
9301 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9303 for_each_possible_cpu(i) {
9305 kfree(tg->cfs_rq[i]);
9308 remove_entity_load_avg(tg->se[i]);
9317 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9319 struct cfs_rq *cfs_rq;
9320 struct sched_entity *se;
9323 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9326 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9330 tg->shares = NICE_0_LOAD;
9332 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9334 for_each_possible_cpu(i) {
9335 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9336 GFP_KERNEL, cpu_to_node(i));
9340 se = kzalloc_node(sizeof(struct sched_entity),
9341 GFP_KERNEL, cpu_to_node(i));
9345 init_cfs_rq(cfs_rq);
9346 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9347 init_entity_runnable_average(se);
9358 void unregister_fair_sched_group(struct task_group *tg, int cpu)
9360 struct rq *rq = cpu_rq(cpu);
9361 unsigned long flags;
9364 * Only empty task groups can be destroyed; so we can speculatively
9365 * check on_list without danger of it being re-added.
9367 if (!tg->cfs_rq[cpu]->on_list)
9370 raw_spin_lock_irqsave(&rq->lock, flags);
9371 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9372 raw_spin_unlock_irqrestore(&rq->lock, flags);
9375 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9376 struct sched_entity *se, int cpu,
9377 struct sched_entity *parent)
9379 struct rq *rq = cpu_rq(cpu);
9383 init_cfs_rq_runtime(cfs_rq);
9385 tg->cfs_rq[cpu] = cfs_rq;
9388 /* se could be NULL for root_task_group */
9393 se->cfs_rq = &rq->cfs;
9396 se->cfs_rq = parent->my_q;
9397 se->depth = parent->depth + 1;
9401 /* guarantee group entities always have weight */
9402 update_load_set(&se->load, NICE_0_LOAD);
9403 se->parent = parent;
9406 static DEFINE_MUTEX(shares_mutex);
9408 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9411 unsigned long flags;
9414 * We can't change the weight of the root cgroup.
9419 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9421 mutex_lock(&shares_mutex);
9422 if (tg->shares == shares)
9425 tg->shares = shares;
9426 for_each_possible_cpu(i) {
9427 struct rq *rq = cpu_rq(i);
9428 struct sched_entity *se;
9431 /* Propagate contribution to hierarchy */
9432 raw_spin_lock_irqsave(&rq->lock, flags);
9434 /* Possible calls to update_curr() need rq clock */
9435 update_rq_clock(rq);
9436 for_each_sched_entity(se)
9437 update_cfs_shares(group_cfs_rq(se));
9438 raw_spin_unlock_irqrestore(&rq->lock, flags);
9442 mutex_unlock(&shares_mutex);
9445 #else /* CONFIG_FAIR_GROUP_SCHED */
9447 void free_fair_sched_group(struct task_group *tg) { }
9449 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9454 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
9456 #endif /* CONFIG_FAIR_GROUP_SCHED */
9459 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
9461 struct sched_entity *se = &task->se;
9462 unsigned int rr_interval = 0;
9465 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
9468 if (rq->cfs.load.weight)
9469 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
9475 * All the scheduling class methods:
9477 const struct sched_class fair_sched_class = {
9478 .next = &idle_sched_class,
9479 .enqueue_task = enqueue_task_fair,
9480 .dequeue_task = dequeue_task_fair,
9481 .yield_task = yield_task_fair,
9482 .yield_to_task = yield_to_task_fair,
9484 .check_preempt_curr = check_preempt_wakeup,
9486 .pick_next_task = pick_next_task_fair,
9487 .put_prev_task = put_prev_task_fair,
9490 .select_task_rq = select_task_rq_fair,
9491 .migrate_task_rq = migrate_task_rq_fair,
9493 .rq_online = rq_online_fair,
9494 .rq_offline = rq_offline_fair,
9496 .task_waking = task_waking_fair,
9497 .task_dead = task_dead_fair,
9498 .set_cpus_allowed = set_cpus_allowed_common,
9501 .set_curr_task = set_curr_task_fair,
9502 .task_tick = task_tick_fair,
9503 .task_fork = task_fork_fair,
9505 .prio_changed = prio_changed_fair,
9506 .switched_from = switched_from_fair,
9507 .switched_to = switched_to_fair,
9509 .get_rr_interval = get_rr_interval_fair,
9511 .update_curr = update_curr_fair,
9513 #ifdef CONFIG_FAIR_GROUP_SCHED
9514 .task_move_group = task_move_group_fair,
9518 #ifdef CONFIG_SCHED_DEBUG
9519 void print_cfs_stats(struct seq_file *m, int cpu)
9521 struct cfs_rq *cfs_rq;
9524 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
9525 print_cfs_rq(m, cpu, cfs_rq);
9529 #ifdef CONFIG_NUMA_BALANCING
9530 void show_numa_stats(struct task_struct *p, struct seq_file *m)
9533 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
9535 for_each_online_node(node) {
9536 if (p->numa_faults) {
9537 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
9538 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
9540 if (p->numa_group) {
9541 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
9542 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
9544 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
9547 #endif /* CONFIG_NUMA_BALANCING */
9548 #endif /* CONFIG_SCHED_DEBUG */
9550 __init void init_sched_fair_class(void)
9553 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9555 #ifdef CONFIG_NO_HZ_COMMON
9556 nohz.next_balance = jiffies;
9557 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
9558 cpu_notifier(sched_ilb_notifier, 0);