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 <pzijlstr@redhat.com>
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>
34 #include <trace/events/sched.h>
39 * Targeted preemption latency for CPU-bound tasks:
40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
42 * NOTE: this latency value is not the same as the concept of
43 * 'timeslice length' - timeslices in CFS are of variable length
44 * and have no persistent notion like in traditional, time-slice
45 * based scheduling concepts.
47 * (to see the precise effective timeslice length of your workload,
48 * run vmstat and monitor the context-switches (cs) field)
50 unsigned int sysctl_sched_latency = 6000000ULL;
51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
54 * The initial- and re-scaling of tunables is configurable
55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
63 = SCHED_TUNABLESCALING_LOG;
66 * Minimal preemption granularity for CPU-bound tasks:
67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
69 unsigned int sysctl_sched_min_granularity = 750000ULL;
70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
75 static unsigned int sched_nr_latency = 8;
78 * After fork, child runs first. If set to 0 (default) then
79 * parent will (try to) run first.
81 unsigned int sysctl_sched_child_runs_first __read_mostly;
84 * SCHED_OTHER wake-up granularity.
85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
87 * This option delays the preemption effects of decoupled workloads
88 * and reduces their over-scheduling. Synchronous workloads will still
89 * have immediate wakeup/sleep latencies.
91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
97 * The exponential sliding window over which load is averaged for shares
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
103 #ifdef CONFIG_CFS_BANDWIDTH
105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106 * each time a cfs_rq requests quota.
108 * Note: in the case that the slice exceeds the runtime remaining (either due
109 * to consumption or the quota being specified to be smaller than the slice)
110 * we will always only issue the remaining available time.
112 * default: 5 msec, units: microseconds
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
136 * Increase the granularity value when there are more CPUs,
137 * because with more CPUs the 'effective latency' as visible
138 * to users decreases. But the relationship is not linear,
139 * so pick a second-best guess by going with the log2 of the
142 * This idea comes from the SD scheduler of Con Kolivas:
144 static unsigned int get_update_sysctl_factor(void)
146 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
149 switch (sysctl_sched_tunable_scaling) {
150 case SCHED_TUNABLESCALING_NONE:
153 case SCHED_TUNABLESCALING_LINEAR:
156 case SCHED_TUNABLESCALING_LOG:
158 factor = 1 + ilog2(cpus);
165 static void update_sysctl(void)
167 unsigned int factor = get_update_sysctl_factor();
169 #define SET_SYSCTL(name) \
170 (sysctl_##name = (factor) * normalized_sysctl_##name)
171 SET_SYSCTL(sched_min_granularity);
172 SET_SYSCTL(sched_latency);
173 SET_SYSCTL(sched_wakeup_granularity);
177 void sched_init_granularity(void)
182 #define WMULT_CONST (~0U)
183 #define WMULT_SHIFT 32
185 static void __update_inv_weight(struct load_weight *lw)
189 if (likely(lw->inv_weight))
192 w = scale_load_down(lw->weight);
194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 else if (unlikely(!w))
197 lw->inv_weight = WMULT_CONST;
199 lw->inv_weight = WMULT_CONST / w;
203 * delta_exec * weight / lw.weight
205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case
208 * we're guaranteed shift stays positive because inv_weight is guaranteed to
209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212 * weight/lw.weight <= 1, and therefore our shift will also be positive.
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
216 u64 fact = scale_load_down(weight);
217 int shift = WMULT_SHIFT;
219 __update_inv_weight(lw);
221 if (unlikely(fact >> 32)) {
228 /* hint to use a 32x32->64 mul */
229 fact = (u64)(u32)fact * lw->inv_weight;
236 return mul_u64_u32_shr(delta_exec, fact, shift);
240 const struct sched_class fair_sched_class;
242 /**************************************************************
243 * CFS operations on generic schedulable entities:
246 #ifdef CONFIG_FAIR_GROUP_SCHED
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se) (!se->my_q)
257 static inline struct task_struct *task_of(struct sched_entity *se)
259 #ifdef CONFIG_SCHED_DEBUG
260 WARN_ON_ONCE(!entity_is_task(se));
262 return container_of(se, struct task_struct, se);
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267 for (; se; se = se->parent)
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
310 if (cfs_rq->on_list) {
311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
324 if (se->cfs_rq == pse->cfs_rq)
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
338 int se_depth, pse_depth;
341 * preemption test can be made between sibling entities who are in the
342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343 * both tasks until we find their ancestors who are siblings of common
347 /* First walk up until both entities are at same depth */
348 se_depth = (*se)->depth;
349 pse_depth = (*pse)->depth;
351 while (se_depth > pse_depth) {
353 *se = parent_entity(*se);
356 while (pse_depth > se_depth) {
358 *pse = parent_entity(*pse);
361 while (!is_same_group(*se, *pse)) {
362 *se = parent_entity(*se);
363 *pse = parent_entity(*pse);
367 #else /* !CONFIG_FAIR_GROUP_SCHED */
369 static inline struct task_struct *task_of(struct sched_entity *se)
371 return container_of(se, struct task_struct, se);
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
376 return container_of(cfs_rq, struct rq, cfs);
379 #define entity_is_task(se) 1
381 #define for_each_sched_entity(se) \
382 for (; se; se = NULL)
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
386 return &task_rq(p)->cfs;
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
391 struct task_struct *p = task_of(se);
392 struct rq *rq = task_rq(p);
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
424 #endif /* CONFIG_FAIR_GROUP_SCHED */
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
429 /**************************************************************
430 * Scheduling class tree data structure manipulation methods:
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
435 s64 delta = (s64)(vruntime - max_vruntime);
437 max_vruntime = vruntime;
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
444 s64 delta = (s64)(vruntime - min_vruntime);
446 min_vruntime = vruntime;
451 static inline int entity_before(struct sched_entity *a,
452 struct sched_entity *b)
454 return (s64)(a->vruntime - b->vruntime) < 0;
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
459 u64 vruntime = cfs_rq->min_vruntime;
462 vruntime = cfs_rq->curr->vruntime;
464 if (cfs_rq->rb_leftmost) {
465 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
470 vruntime = se->vruntime;
472 vruntime = min_vruntime(vruntime, se->vruntime);
475 /* ensure we never gain time by being placed backwards. */
476 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
479 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
484 * Enqueue an entity into the rb-tree:
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489 struct rb_node *parent = NULL;
490 struct sched_entity *entry;
494 * Find the right place in the rbtree:
498 entry = rb_entry(parent, struct sched_entity, run_node);
500 * We dont care about collisions. Nodes with
501 * the same key stay together.
503 if (entity_before(se, entry)) {
504 link = &parent->rb_left;
506 link = &parent->rb_right;
512 * Maintain a cache of leftmost tree entries (it is frequently
516 cfs_rq->rb_leftmost = &se->run_node;
518 rb_link_node(&se->run_node, parent, link);
519 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 if (cfs_rq->rb_leftmost == &se->run_node) {
525 struct rb_node *next_node;
527 next_node = rb_next(&se->run_node);
528 cfs_rq->rb_leftmost = next_node;
531 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
536 struct rb_node *left = cfs_rq->rb_leftmost;
541 return rb_entry(left, struct sched_entity, run_node);
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
546 struct rb_node *next = rb_next(&se->run_node);
551 return rb_entry(next, struct sched_entity, run_node);
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
557 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
562 return rb_entry(last, struct sched_entity, run_node);
565 /**************************************************************
566 * Scheduling class statistics methods:
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp,
573 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574 unsigned int factor = get_update_sysctl_factor();
579 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580 sysctl_sched_min_granularity);
582 #define WRT_SYSCTL(name) \
583 (normalized_sysctl_##name = sysctl_##name / (factor))
584 WRT_SYSCTL(sched_min_granularity);
585 WRT_SYSCTL(sched_latency);
586 WRT_SYSCTL(sched_wakeup_granularity);
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
598 if (unlikely(se->load.weight != NICE_0_LOAD))
599 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
605 * The idea is to set a period in which each task runs once.
607 * When there are too many tasks (sched_nr_latency) we have to stretch
608 * this period because otherwise the slices get too small.
610 * p = (nr <= nl) ? l : l*nr/nl
612 static u64 __sched_period(unsigned long nr_running)
614 if (unlikely(nr_running > sched_nr_latency))
615 return nr_running * sysctl_sched_min_granularity;
617 return sysctl_sched_latency;
621 * We calculate the wall-time slice from the period by taking a part
622 * proportional to the weight.
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
630 for_each_sched_entity(se) {
631 struct load_weight *load;
632 struct load_weight lw;
634 cfs_rq = cfs_rq_of(se);
635 load = &cfs_rq->load;
637 if (unlikely(!se->on_rq)) {
640 update_load_add(&lw, se->load.weight);
643 slice = __calc_delta(slice, se->load.weight, load);
649 * We calculate the vruntime slice of a to-be-inserted task.
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 return calc_delta_fair(sched_slice(cfs_rq, se), se);
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
663 * We choose a half-life close to 1 scheduling period.
664 * Note: The tables below are dependent on this value.
666 #define LOAD_AVG_PERIOD 32
667 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
668 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
670 /* Give new task start runnable values to heavy its load in infant time */
671 void init_task_runnable_average(struct task_struct *p)
673 struct sched_avg *sa = &p->se.avg;
675 sa->last_update_time = 0;
677 * sched_avg's period_contrib should be strictly less then 1024, so
678 * we give it 1023 to make sure it is almost a period (1024us), and
679 * will definitely be update (after enqueue).
681 sa->period_contrib = 1023;
682 sa->load_avg = scale_load_down(p->se.load.weight);
683 sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
684 sa->util_avg = scale_load_down(SCHED_LOAD_SCALE);
685 sa->util_sum = LOAD_AVG_MAX;
686 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
689 void init_task_runnable_average(struct task_struct *p)
695 * Update the current task's runtime statistics.
697 static void update_curr(struct cfs_rq *cfs_rq)
699 struct sched_entity *curr = cfs_rq->curr;
700 u64 now = rq_clock_task(rq_of(cfs_rq));
706 delta_exec = now - curr->exec_start;
707 if (unlikely((s64)delta_exec <= 0))
710 curr->exec_start = now;
712 schedstat_set(curr->statistics.exec_max,
713 max(delta_exec, curr->statistics.exec_max));
715 curr->sum_exec_runtime += delta_exec;
716 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 curr->vruntime += calc_delta_fair(delta_exec, curr);
719 update_min_vruntime(cfs_rq);
721 if (entity_is_task(curr)) {
722 struct task_struct *curtask = task_of(curr);
724 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
725 cpuacct_charge(curtask, delta_exec);
726 account_group_exec_runtime(curtask, delta_exec);
729 account_cfs_rq_runtime(cfs_rq, delta_exec);
732 static void update_curr_fair(struct rq *rq)
734 update_curr(cfs_rq_of(&rq->curr->se));
738 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
740 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
744 * Task is being enqueued - update stats:
746 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
749 * Are we enqueueing a waiting task? (for current tasks
750 * a dequeue/enqueue event is a NOP)
752 if (se != cfs_rq->curr)
753 update_stats_wait_start(cfs_rq, se);
757 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
760 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
761 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
762 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
764 #ifdef CONFIG_SCHEDSTATS
765 if (entity_is_task(se)) {
766 trace_sched_stat_wait(task_of(se),
767 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
770 schedstat_set(se->statistics.wait_start, 0);
774 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
777 * Mark the end of the wait period if dequeueing a
780 if (se != cfs_rq->curr)
781 update_stats_wait_end(cfs_rq, se);
785 * We are picking a new current task - update its stats:
788 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
791 * We are starting a new run period:
793 se->exec_start = rq_clock_task(rq_of(cfs_rq));
796 /**************************************************
797 * Scheduling class queueing methods:
800 #ifdef CONFIG_NUMA_BALANCING
802 * Approximate time to scan a full NUMA task in ms. The task scan period is
803 * calculated based on the tasks virtual memory size and
804 * numa_balancing_scan_size.
806 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
807 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
809 /* Portion of address space to scan in MB */
810 unsigned int sysctl_numa_balancing_scan_size = 256;
812 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
813 unsigned int sysctl_numa_balancing_scan_delay = 1000;
815 static unsigned int task_nr_scan_windows(struct task_struct *p)
817 unsigned long rss = 0;
818 unsigned long nr_scan_pages;
821 * Calculations based on RSS as non-present and empty pages are skipped
822 * by the PTE scanner and NUMA hinting faults should be trapped based
825 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
826 rss = get_mm_rss(p->mm);
830 rss = round_up(rss, nr_scan_pages);
831 return rss / nr_scan_pages;
834 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
835 #define MAX_SCAN_WINDOW 2560
837 static unsigned int task_scan_min(struct task_struct *p)
839 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
840 unsigned int scan, floor;
841 unsigned int windows = 1;
843 if (scan_size < MAX_SCAN_WINDOW)
844 windows = MAX_SCAN_WINDOW / scan_size;
845 floor = 1000 / windows;
847 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
848 return max_t(unsigned int, floor, scan);
851 static unsigned int task_scan_max(struct task_struct *p)
853 unsigned int smin = task_scan_min(p);
856 /* Watch for min being lower than max due to floor calculations */
857 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
858 return max(smin, smax);
861 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
863 rq->nr_numa_running += (p->numa_preferred_nid != -1);
864 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
867 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
869 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
870 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
876 spinlock_t lock; /* nr_tasks, tasks */
881 nodemask_t active_nodes;
882 unsigned long total_faults;
884 * Faults_cpu is used to decide whether memory should move
885 * towards the CPU. As a consequence, these stats are weighted
886 * more by CPU use than by memory faults.
888 unsigned long *faults_cpu;
889 unsigned long faults[0];
892 /* Shared or private faults. */
893 #define NR_NUMA_HINT_FAULT_TYPES 2
895 /* Memory and CPU locality */
896 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
898 /* Averaged statistics, and temporary buffers. */
899 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
901 pid_t task_numa_group_id(struct task_struct *p)
903 return p->numa_group ? p->numa_group->gid : 0;
907 * The averaged statistics, shared & private, memory & cpu,
908 * occupy the first half of the array. The second half of the
909 * array is for current counters, which are averaged into the
910 * first set by task_numa_placement.
912 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
914 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
917 static inline unsigned long task_faults(struct task_struct *p, int nid)
922 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
923 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
926 static inline unsigned long group_faults(struct task_struct *p, int nid)
931 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
932 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
935 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
937 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
938 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
941 /* Handle placement on systems where not all nodes are directly connected. */
942 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
943 int maxdist, bool task)
945 unsigned long score = 0;
949 * All nodes are directly connected, and the same distance
950 * from each other. No need for fancy placement algorithms.
952 if (sched_numa_topology_type == NUMA_DIRECT)
956 * This code is called for each node, introducing N^2 complexity,
957 * which should be ok given the number of nodes rarely exceeds 8.
959 for_each_online_node(node) {
960 unsigned long faults;
961 int dist = node_distance(nid, node);
964 * The furthest away nodes in the system are not interesting
965 * for placement; nid was already counted.
967 if (dist == sched_max_numa_distance || node == nid)
971 * On systems with a backplane NUMA topology, compare groups
972 * of nodes, and move tasks towards the group with the most
973 * memory accesses. When comparing two nodes at distance
974 * "hoplimit", only nodes closer by than "hoplimit" are part
975 * of each group. Skip other nodes.
977 if (sched_numa_topology_type == NUMA_BACKPLANE &&
981 /* Add up the faults from nearby nodes. */
983 faults = task_faults(p, node);
985 faults = group_faults(p, node);
988 * On systems with a glueless mesh NUMA topology, there are
989 * no fixed "groups of nodes". Instead, nodes that are not
990 * directly connected bounce traffic through intermediate
991 * nodes; a numa_group can occupy any set of nodes.
992 * The further away a node is, the less the faults count.
993 * This seems to result in good task placement.
995 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
996 faults *= (sched_max_numa_distance - dist);
997 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1007 * These return the fraction of accesses done by a particular task, or
1008 * task group, on a particular numa node. The group weight is given a
1009 * larger multiplier, in order to group tasks together that are almost
1010 * evenly spread out between numa nodes.
1012 static inline unsigned long task_weight(struct task_struct *p, int nid,
1015 unsigned long faults, total_faults;
1017 if (!p->numa_faults)
1020 total_faults = p->total_numa_faults;
1025 faults = task_faults(p, nid);
1026 faults += score_nearby_nodes(p, nid, dist, true);
1028 return 1000 * faults / total_faults;
1031 static inline unsigned long group_weight(struct task_struct *p, int nid,
1034 unsigned long faults, total_faults;
1039 total_faults = p->numa_group->total_faults;
1044 faults = group_faults(p, nid);
1045 faults += score_nearby_nodes(p, nid, dist, false);
1047 return 1000 * faults / total_faults;
1050 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1051 int src_nid, int dst_cpu)
1053 struct numa_group *ng = p->numa_group;
1054 int dst_nid = cpu_to_node(dst_cpu);
1055 int last_cpupid, this_cpupid;
1057 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1060 * Multi-stage node selection is used in conjunction with a periodic
1061 * migration fault to build a temporal task<->page relation. By using
1062 * a two-stage filter we remove short/unlikely relations.
1064 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1065 * a task's usage of a particular page (n_p) per total usage of this
1066 * page (n_t) (in a given time-span) to a probability.
1068 * Our periodic faults will sample this probability and getting the
1069 * same result twice in a row, given these samples are fully
1070 * independent, is then given by P(n)^2, provided our sample period
1071 * is sufficiently short compared to the usage pattern.
1073 * This quadric squishes small probabilities, making it less likely we
1074 * act on an unlikely task<->page relation.
1076 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1077 if (!cpupid_pid_unset(last_cpupid) &&
1078 cpupid_to_nid(last_cpupid) != dst_nid)
1081 /* Always allow migrate on private faults */
1082 if (cpupid_match_pid(p, last_cpupid))
1085 /* A shared fault, but p->numa_group has not been set up yet. */
1090 * Do not migrate if the destination is not a node that
1091 * is actively used by this numa group.
1093 if (!node_isset(dst_nid, ng->active_nodes))
1097 * Source is a node that is not actively used by this
1098 * numa group, while the destination is. Migrate.
1100 if (!node_isset(src_nid, ng->active_nodes))
1104 * Both source and destination are nodes in active
1105 * use by this numa group. Maximize memory bandwidth
1106 * by migrating from more heavily used groups, to less
1107 * heavily used ones, spreading the load around.
1108 * Use a 1/4 hysteresis to avoid spurious page movement.
1110 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4);
1113 static unsigned long weighted_cpuload(const int cpu);
1114 static unsigned long source_load(int cpu, int type);
1115 static unsigned long target_load(int cpu, int type);
1116 static unsigned long capacity_of(int cpu);
1117 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1119 /* Cached statistics for all CPUs within a node */
1121 unsigned long nr_running;
1124 /* Total compute capacity of CPUs on a node */
1125 unsigned long compute_capacity;
1127 /* Approximate capacity in terms of runnable tasks on a node */
1128 unsigned long task_capacity;
1129 int has_free_capacity;
1133 * XXX borrowed from update_sg_lb_stats
1135 static void update_numa_stats(struct numa_stats *ns, int nid)
1137 int smt, cpu, cpus = 0;
1138 unsigned long capacity;
1140 memset(ns, 0, sizeof(*ns));
1141 for_each_cpu(cpu, cpumask_of_node(nid)) {
1142 struct rq *rq = cpu_rq(cpu);
1144 ns->nr_running += rq->nr_running;
1145 ns->load += weighted_cpuload(cpu);
1146 ns->compute_capacity += capacity_of(cpu);
1152 * If we raced with hotplug and there are no CPUs left in our mask
1153 * the @ns structure is NULL'ed and task_numa_compare() will
1154 * not find this node attractive.
1156 * We'll either bail at !has_free_capacity, or we'll detect a huge
1157 * imbalance and bail there.
1162 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1163 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1164 capacity = cpus / smt; /* cores */
1166 ns->task_capacity = min_t(unsigned, capacity,
1167 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1168 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1171 struct task_numa_env {
1172 struct task_struct *p;
1174 int src_cpu, src_nid;
1175 int dst_cpu, dst_nid;
1177 struct numa_stats src_stats, dst_stats;
1182 struct task_struct *best_task;
1187 static void task_numa_assign(struct task_numa_env *env,
1188 struct task_struct *p, long imp)
1191 put_task_struct(env->best_task);
1196 env->best_imp = imp;
1197 env->best_cpu = env->dst_cpu;
1200 static bool load_too_imbalanced(long src_load, long dst_load,
1201 struct task_numa_env *env)
1204 long orig_src_load, orig_dst_load;
1205 long src_capacity, dst_capacity;
1208 * The load is corrected for the CPU capacity available on each node.
1211 * ------------ vs ---------
1212 * src_capacity dst_capacity
1214 src_capacity = env->src_stats.compute_capacity;
1215 dst_capacity = env->dst_stats.compute_capacity;
1217 /* We care about the slope of the imbalance, not the direction. */
1218 if (dst_load < src_load)
1219 swap(dst_load, src_load);
1221 /* Is the difference below the threshold? */
1222 imb = dst_load * src_capacity * 100 -
1223 src_load * dst_capacity * env->imbalance_pct;
1228 * The imbalance is above the allowed threshold.
1229 * Compare it with the old imbalance.
1231 orig_src_load = env->src_stats.load;
1232 orig_dst_load = env->dst_stats.load;
1234 if (orig_dst_load < orig_src_load)
1235 swap(orig_dst_load, orig_src_load);
1237 old_imb = orig_dst_load * src_capacity * 100 -
1238 orig_src_load * dst_capacity * env->imbalance_pct;
1240 /* Would this change make things worse? */
1241 return (imb > old_imb);
1245 * This checks if the overall compute and NUMA accesses of the system would
1246 * be improved if the source tasks was migrated to the target dst_cpu taking
1247 * into account that it might be best if task running on the dst_cpu should
1248 * be exchanged with the source task
1250 static void task_numa_compare(struct task_numa_env *env,
1251 long taskimp, long groupimp)
1253 struct rq *src_rq = cpu_rq(env->src_cpu);
1254 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1255 struct task_struct *cur;
1256 long src_load, dst_load;
1258 long imp = env->p->numa_group ? groupimp : taskimp;
1260 int dist = env->dist;
1264 raw_spin_lock_irq(&dst_rq->lock);
1267 * No need to move the exiting task, and this ensures that ->curr
1268 * wasn't reaped and thus get_task_struct() in task_numa_assign()
1269 * is safe under RCU read lock.
1270 * Note that rcu_read_lock() itself can't protect from the final
1271 * put_task_struct() after the last schedule().
1273 if ((cur->flags & PF_EXITING) || is_idle_task(cur))
1275 raw_spin_unlock_irq(&dst_rq->lock);
1278 * Because we have preemption enabled we can get migrated around and
1279 * end try selecting ourselves (current == env->p) as a swap candidate.
1285 * "imp" is the fault differential for the source task between the
1286 * source and destination node. Calculate the total differential for
1287 * the source task and potential destination task. The more negative
1288 * the value is, the more rmeote accesses that would be expected to
1289 * be incurred if the tasks were swapped.
1292 /* Skip this swap candidate if cannot move to the source cpu */
1293 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1297 * If dst and source tasks are in the same NUMA group, or not
1298 * in any group then look only at task weights.
1300 if (cur->numa_group == env->p->numa_group) {
1301 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1302 task_weight(cur, env->dst_nid, dist);
1304 * Add some hysteresis to prevent swapping the
1305 * tasks within a group over tiny differences.
1307 if (cur->numa_group)
1311 * Compare the group weights. If a task is all by
1312 * itself (not part of a group), use the task weight
1315 if (cur->numa_group)
1316 imp += group_weight(cur, env->src_nid, dist) -
1317 group_weight(cur, env->dst_nid, dist);
1319 imp += task_weight(cur, env->src_nid, dist) -
1320 task_weight(cur, env->dst_nid, dist);
1324 if (imp <= env->best_imp && moveimp <= env->best_imp)
1328 /* Is there capacity at our destination? */
1329 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1330 !env->dst_stats.has_free_capacity)
1336 /* Balance doesn't matter much if we're running a task per cpu */
1337 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1338 dst_rq->nr_running == 1)
1342 * In the overloaded case, try and keep the load balanced.
1345 load = task_h_load(env->p);
1346 dst_load = env->dst_stats.load + load;
1347 src_load = env->src_stats.load - load;
1349 if (moveimp > imp && moveimp > env->best_imp) {
1351 * If the improvement from just moving env->p direction is
1352 * better than swapping tasks around, check if a move is
1353 * possible. Store a slightly smaller score than moveimp,
1354 * so an actually idle CPU will win.
1356 if (!load_too_imbalanced(src_load, dst_load, env)) {
1363 if (imp <= env->best_imp)
1367 load = task_h_load(cur);
1372 if (load_too_imbalanced(src_load, dst_load, env))
1376 * One idle CPU per node is evaluated for a task numa move.
1377 * Call select_idle_sibling to maybe find a better one.
1380 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1383 task_numa_assign(env, cur, imp);
1388 static void task_numa_find_cpu(struct task_numa_env *env,
1389 long taskimp, long groupimp)
1393 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1394 /* Skip this CPU if the source task cannot migrate */
1395 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1399 task_numa_compare(env, taskimp, groupimp);
1403 /* Only move tasks to a NUMA node less busy than the current node. */
1404 static bool numa_has_capacity(struct task_numa_env *env)
1406 struct numa_stats *src = &env->src_stats;
1407 struct numa_stats *dst = &env->dst_stats;
1409 if (src->has_free_capacity && !dst->has_free_capacity)
1413 * Only consider a task move if the source has a higher load
1414 * than the destination, corrected for CPU capacity on each node.
1416 * src->load dst->load
1417 * --------------------- vs ---------------------
1418 * src->compute_capacity dst->compute_capacity
1420 if (src->load * dst->compute_capacity * env->imbalance_pct >
1422 dst->load * src->compute_capacity * 100)
1428 static int task_numa_migrate(struct task_struct *p)
1430 struct task_numa_env env = {
1433 .src_cpu = task_cpu(p),
1434 .src_nid = task_node(p),
1436 .imbalance_pct = 112,
1442 struct sched_domain *sd;
1443 unsigned long taskweight, groupweight;
1445 long taskimp, groupimp;
1448 * Pick the lowest SD_NUMA domain, as that would have the smallest
1449 * imbalance and would be the first to start moving tasks about.
1451 * And we want to avoid any moving of tasks about, as that would create
1452 * random movement of tasks -- counter the numa conditions we're trying
1456 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1458 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1462 * Cpusets can break the scheduler domain tree into smaller
1463 * balance domains, some of which do not cross NUMA boundaries.
1464 * Tasks that are "trapped" in such domains cannot be migrated
1465 * elsewhere, so there is no point in (re)trying.
1467 if (unlikely(!sd)) {
1468 p->numa_preferred_nid = task_node(p);
1472 env.dst_nid = p->numa_preferred_nid;
1473 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1474 taskweight = task_weight(p, env.src_nid, dist);
1475 groupweight = group_weight(p, env.src_nid, dist);
1476 update_numa_stats(&env.src_stats, env.src_nid);
1477 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1478 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1479 update_numa_stats(&env.dst_stats, env.dst_nid);
1481 /* Try to find a spot on the preferred nid. */
1482 if (numa_has_capacity(&env))
1483 task_numa_find_cpu(&env, taskimp, groupimp);
1486 * Look at other nodes in these cases:
1487 * - there is no space available on the preferred_nid
1488 * - the task is part of a numa_group that is interleaved across
1489 * multiple NUMA nodes; in order to better consolidate the group,
1490 * we need to check other locations.
1492 if (env.best_cpu == -1 || (p->numa_group &&
1493 nodes_weight(p->numa_group->active_nodes) > 1)) {
1494 for_each_online_node(nid) {
1495 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1498 dist = node_distance(env.src_nid, env.dst_nid);
1499 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1501 taskweight = task_weight(p, env.src_nid, dist);
1502 groupweight = group_weight(p, env.src_nid, dist);
1505 /* Only consider nodes where both task and groups benefit */
1506 taskimp = task_weight(p, nid, dist) - taskweight;
1507 groupimp = group_weight(p, nid, dist) - groupweight;
1508 if (taskimp < 0 && groupimp < 0)
1513 update_numa_stats(&env.dst_stats, env.dst_nid);
1514 if (numa_has_capacity(&env))
1515 task_numa_find_cpu(&env, taskimp, groupimp);
1520 * If the task is part of a workload that spans multiple NUMA nodes,
1521 * and is migrating into one of the workload's active nodes, remember
1522 * this node as the task's preferred numa node, so the workload can
1524 * A task that migrated to a second choice node will be better off
1525 * trying for a better one later. Do not set the preferred node here.
1527 if (p->numa_group) {
1528 if (env.best_cpu == -1)
1533 if (node_isset(nid, p->numa_group->active_nodes))
1534 sched_setnuma(p, env.dst_nid);
1537 /* No better CPU than the current one was found. */
1538 if (env.best_cpu == -1)
1542 * Reset the scan period if the task is being rescheduled on an
1543 * alternative node to recheck if the tasks is now properly placed.
1545 p->numa_scan_period = task_scan_min(p);
1547 if (env.best_task == NULL) {
1548 ret = migrate_task_to(p, env.best_cpu);
1550 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1554 ret = migrate_swap(p, env.best_task);
1556 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1557 put_task_struct(env.best_task);
1561 /* Attempt to migrate a task to a CPU on the preferred node. */
1562 static void numa_migrate_preferred(struct task_struct *p)
1564 unsigned long interval = HZ;
1566 /* This task has no NUMA fault statistics yet */
1567 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1570 /* Periodically retry migrating the task to the preferred node */
1571 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1572 p->numa_migrate_retry = jiffies + interval;
1574 /* Success if task is already running on preferred CPU */
1575 if (task_node(p) == p->numa_preferred_nid)
1578 /* Otherwise, try migrate to a CPU on the preferred node */
1579 task_numa_migrate(p);
1583 * Find the nodes on which the workload is actively running. We do this by
1584 * tracking the nodes from which NUMA hinting faults are triggered. This can
1585 * be different from the set of nodes where the workload's memory is currently
1588 * The bitmask is used to make smarter decisions on when to do NUMA page
1589 * migrations, To prevent flip-flopping, and excessive page migrations, nodes
1590 * are added when they cause over 6/16 of the maximum number of faults, but
1591 * only removed when they drop below 3/16.
1593 static void update_numa_active_node_mask(struct numa_group *numa_group)
1595 unsigned long faults, max_faults = 0;
1598 for_each_online_node(nid) {
1599 faults = group_faults_cpu(numa_group, nid);
1600 if (faults > max_faults)
1601 max_faults = faults;
1604 for_each_online_node(nid) {
1605 faults = group_faults_cpu(numa_group, nid);
1606 if (!node_isset(nid, numa_group->active_nodes)) {
1607 if (faults > max_faults * 6 / 16)
1608 node_set(nid, numa_group->active_nodes);
1609 } else if (faults < max_faults * 3 / 16)
1610 node_clear(nid, numa_group->active_nodes);
1615 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1616 * increments. The more local the fault statistics are, the higher the scan
1617 * period will be for the next scan window. If local/(local+remote) ratio is
1618 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1619 * the scan period will decrease. Aim for 70% local accesses.
1621 #define NUMA_PERIOD_SLOTS 10
1622 #define NUMA_PERIOD_THRESHOLD 7
1625 * Increase the scan period (slow down scanning) if the majority of
1626 * our memory is already on our local node, or if the majority of
1627 * the page accesses are shared with other processes.
1628 * Otherwise, decrease the scan period.
1630 static void update_task_scan_period(struct task_struct *p,
1631 unsigned long shared, unsigned long private)
1633 unsigned int period_slot;
1637 unsigned long remote = p->numa_faults_locality[0];
1638 unsigned long local = p->numa_faults_locality[1];
1641 * If there were no record hinting faults then either the task is
1642 * completely idle or all activity is areas that are not of interest
1643 * to automatic numa balancing. Related to that, if there were failed
1644 * migration then it implies we are migrating too quickly or the local
1645 * node is overloaded. In either case, scan slower
1647 if (local + shared == 0 || p->numa_faults_locality[2]) {
1648 p->numa_scan_period = min(p->numa_scan_period_max,
1649 p->numa_scan_period << 1);
1651 p->mm->numa_next_scan = jiffies +
1652 msecs_to_jiffies(p->numa_scan_period);
1658 * Prepare to scale scan period relative to the current period.
1659 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1660 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1661 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1663 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1664 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1665 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1666 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1669 diff = slot * period_slot;
1671 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1674 * Scale scan rate increases based on sharing. There is an
1675 * inverse relationship between the degree of sharing and
1676 * the adjustment made to the scanning period. Broadly
1677 * speaking the intent is that there is little point
1678 * scanning faster if shared accesses dominate as it may
1679 * simply bounce migrations uselessly
1681 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1682 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1685 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1686 task_scan_min(p), task_scan_max(p));
1687 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1691 * Get the fraction of time the task has been running since the last
1692 * NUMA placement cycle. The scheduler keeps similar statistics, but
1693 * decays those on a 32ms period, which is orders of magnitude off
1694 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1695 * stats only if the task is so new there are no NUMA statistics yet.
1697 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1699 u64 runtime, delta, now;
1700 /* Use the start of this time slice to avoid calculations. */
1701 now = p->se.exec_start;
1702 runtime = p->se.sum_exec_runtime;
1704 if (p->last_task_numa_placement) {
1705 delta = runtime - p->last_sum_exec_runtime;
1706 *period = now - p->last_task_numa_placement;
1708 delta = p->se.avg.load_sum / p->se.load.weight;
1709 *period = LOAD_AVG_MAX;
1712 p->last_sum_exec_runtime = runtime;
1713 p->last_task_numa_placement = now;
1719 * Determine the preferred nid for a task in a numa_group. This needs to
1720 * be done in a way that produces consistent results with group_weight,
1721 * otherwise workloads might not converge.
1723 static int preferred_group_nid(struct task_struct *p, int nid)
1728 /* Direct connections between all NUMA nodes. */
1729 if (sched_numa_topology_type == NUMA_DIRECT)
1733 * On a system with glueless mesh NUMA topology, group_weight
1734 * scores nodes according to the number of NUMA hinting faults on
1735 * both the node itself, and on nearby nodes.
1737 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1738 unsigned long score, max_score = 0;
1739 int node, max_node = nid;
1741 dist = sched_max_numa_distance;
1743 for_each_online_node(node) {
1744 score = group_weight(p, node, dist);
1745 if (score > max_score) {
1754 * Finding the preferred nid in a system with NUMA backplane
1755 * interconnect topology is more involved. The goal is to locate
1756 * tasks from numa_groups near each other in the system, and
1757 * untangle workloads from different sides of the system. This requires
1758 * searching down the hierarchy of node groups, recursively searching
1759 * inside the highest scoring group of nodes. The nodemask tricks
1760 * keep the complexity of the search down.
1762 nodes = node_online_map;
1763 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1764 unsigned long max_faults = 0;
1765 nodemask_t max_group = NODE_MASK_NONE;
1768 /* Are there nodes at this distance from each other? */
1769 if (!find_numa_distance(dist))
1772 for_each_node_mask(a, nodes) {
1773 unsigned long faults = 0;
1774 nodemask_t this_group;
1775 nodes_clear(this_group);
1777 /* Sum group's NUMA faults; includes a==b case. */
1778 for_each_node_mask(b, nodes) {
1779 if (node_distance(a, b) < dist) {
1780 faults += group_faults(p, b);
1781 node_set(b, this_group);
1782 node_clear(b, nodes);
1786 /* Remember the top group. */
1787 if (faults > max_faults) {
1788 max_faults = faults;
1789 max_group = this_group;
1791 * subtle: at the smallest distance there is
1792 * just one node left in each "group", the
1793 * winner is the preferred nid.
1798 /* Next round, evaluate the nodes within max_group. */
1806 static void task_numa_placement(struct task_struct *p)
1808 int seq, nid, max_nid = -1, max_group_nid = -1;
1809 unsigned long max_faults = 0, max_group_faults = 0;
1810 unsigned long fault_types[2] = { 0, 0 };
1811 unsigned long total_faults;
1812 u64 runtime, period;
1813 spinlock_t *group_lock = NULL;
1816 * The p->mm->numa_scan_seq field gets updated without
1817 * exclusive access. Use READ_ONCE() here to ensure
1818 * that the field is read in a single access:
1820 seq = READ_ONCE(p->mm->numa_scan_seq);
1821 if (p->numa_scan_seq == seq)
1823 p->numa_scan_seq = seq;
1824 p->numa_scan_period_max = task_scan_max(p);
1826 total_faults = p->numa_faults_locality[0] +
1827 p->numa_faults_locality[1];
1828 runtime = numa_get_avg_runtime(p, &period);
1830 /* If the task is part of a group prevent parallel updates to group stats */
1831 if (p->numa_group) {
1832 group_lock = &p->numa_group->lock;
1833 spin_lock_irq(group_lock);
1836 /* Find the node with the highest number of faults */
1837 for_each_online_node(nid) {
1838 /* Keep track of the offsets in numa_faults array */
1839 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1840 unsigned long faults = 0, group_faults = 0;
1843 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1844 long diff, f_diff, f_weight;
1846 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1847 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1848 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1849 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1851 /* Decay existing window, copy faults since last scan */
1852 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1853 fault_types[priv] += p->numa_faults[membuf_idx];
1854 p->numa_faults[membuf_idx] = 0;
1857 * Normalize the faults_from, so all tasks in a group
1858 * count according to CPU use, instead of by the raw
1859 * number of faults. Tasks with little runtime have
1860 * little over-all impact on throughput, and thus their
1861 * faults are less important.
1863 f_weight = div64_u64(runtime << 16, period + 1);
1864 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1866 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
1867 p->numa_faults[cpubuf_idx] = 0;
1869 p->numa_faults[mem_idx] += diff;
1870 p->numa_faults[cpu_idx] += f_diff;
1871 faults += p->numa_faults[mem_idx];
1872 p->total_numa_faults += diff;
1873 if (p->numa_group) {
1875 * safe because we can only change our own group
1877 * mem_idx represents the offset for a given
1878 * nid and priv in a specific region because it
1879 * is at the beginning of the numa_faults array.
1881 p->numa_group->faults[mem_idx] += diff;
1882 p->numa_group->faults_cpu[mem_idx] += f_diff;
1883 p->numa_group->total_faults += diff;
1884 group_faults += p->numa_group->faults[mem_idx];
1888 if (faults > max_faults) {
1889 max_faults = faults;
1893 if (group_faults > max_group_faults) {
1894 max_group_faults = group_faults;
1895 max_group_nid = nid;
1899 update_task_scan_period(p, fault_types[0], fault_types[1]);
1901 if (p->numa_group) {
1902 update_numa_active_node_mask(p->numa_group);
1903 spin_unlock_irq(group_lock);
1904 max_nid = preferred_group_nid(p, max_group_nid);
1908 /* Set the new preferred node */
1909 if (max_nid != p->numa_preferred_nid)
1910 sched_setnuma(p, max_nid);
1912 if (task_node(p) != p->numa_preferred_nid)
1913 numa_migrate_preferred(p);
1917 static inline int get_numa_group(struct numa_group *grp)
1919 return atomic_inc_not_zero(&grp->refcount);
1922 static inline void put_numa_group(struct numa_group *grp)
1924 if (atomic_dec_and_test(&grp->refcount))
1925 kfree_rcu(grp, rcu);
1928 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1931 struct numa_group *grp, *my_grp;
1932 struct task_struct *tsk;
1934 int cpu = cpupid_to_cpu(cpupid);
1937 if (unlikely(!p->numa_group)) {
1938 unsigned int size = sizeof(struct numa_group) +
1939 4*nr_node_ids*sizeof(unsigned long);
1941 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1945 atomic_set(&grp->refcount, 1);
1946 spin_lock_init(&grp->lock);
1948 /* Second half of the array tracks nids where faults happen */
1949 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
1952 node_set(task_node(current), grp->active_nodes);
1954 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
1955 grp->faults[i] = p->numa_faults[i];
1957 grp->total_faults = p->total_numa_faults;
1960 rcu_assign_pointer(p->numa_group, grp);
1964 tsk = READ_ONCE(cpu_rq(cpu)->curr);
1966 if (!cpupid_match_pid(tsk, cpupid))
1969 grp = rcu_dereference(tsk->numa_group);
1973 my_grp = p->numa_group;
1978 * Only join the other group if its bigger; if we're the bigger group,
1979 * the other task will join us.
1981 if (my_grp->nr_tasks > grp->nr_tasks)
1985 * Tie-break on the grp address.
1987 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1990 /* Always join threads in the same process. */
1991 if (tsk->mm == current->mm)
1994 /* Simple filter to avoid false positives due to PID collisions */
1995 if (flags & TNF_SHARED)
1998 /* Update priv based on whether false sharing was detected */
2001 if (join && !get_numa_group(grp))
2009 BUG_ON(irqs_disabled());
2010 double_lock_irq(&my_grp->lock, &grp->lock);
2012 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2013 my_grp->faults[i] -= p->numa_faults[i];
2014 grp->faults[i] += p->numa_faults[i];
2016 my_grp->total_faults -= p->total_numa_faults;
2017 grp->total_faults += p->total_numa_faults;
2022 spin_unlock(&my_grp->lock);
2023 spin_unlock_irq(&grp->lock);
2025 rcu_assign_pointer(p->numa_group, grp);
2027 put_numa_group(my_grp);
2035 void task_numa_free(struct task_struct *p)
2037 struct numa_group *grp = p->numa_group;
2038 void *numa_faults = p->numa_faults;
2039 unsigned long flags;
2043 spin_lock_irqsave(&grp->lock, flags);
2044 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2045 grp->faults[i] -= p->numa_faults[i];
2046 grp->total_faults -= p->total_numa_faults;
2049 spin_unlock_irqrestore(&grp->lock, flags);
2050 RCU_INIT_POINTER(p->numa_group, NULL);
2051 put_numa_group(grp);
2054 p->numa_faults = NULL;
2059 * Got a PROT_NONE fault for a page on @node.
2061 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2063 struct task_struct *p = current;
2064 bool migrated = flags & TNF_MIGRATED;
2065 int cpu_node = task_node(current);
2066 int local = !!(flags & TNF_FAULT_LOCAL);
2069 if (!numabalancing_enabled)
2072 /* for example, ksmd faulting in a user's mm */
2076 /* Allocate buffer to track faults on a per-node basis */
2077 if (unlikely(!p->numa_faults)) {
2078 int size = sizeof(*p->numa_faults) *
2079 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2081 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2082 if (!p->numa_faults)
2085 p->total_numa_faults = 0;
2086 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2090 * First accesses are treated as private, otherwise consider accesses
2091 * to be private if the accessing pid has not changed
2093 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2096 priv = cpupid_match_pid(p, last_cpupid);
2097 if (!priv && !(flags & TNF_NO_GROUP))
2098 task_numa_group(p, last_cpupid, flags, &priv);
2102 * If a workload spans multiple NUMA nodes, a shared fault that
2103 * occurs wholly within the set of nodes that the workload is
2104 * actively using should be counted as local. This allows the
2105 * scan rate to slow down when a workload has settled down.
2107 if (!priv && !local && p->numa_group &&
2108 node_isset(cpu_node, p->numa_group->active_nodes) &&
2109 node_isset(mem_node, p->numa_group->active_nodes))
2112 task_numa_placement(p);
2115 * Retry task to preferred node migration periodically, in case it
2116 * case it previously failed, or the scheduler moved us.
2118 if (time_after(jiffies, p->numa_migrate_retry))
2119 numa_migrate_preferred(p);
2122 p->numa_pages_migrated += pages;
2123 if (flags & TNF_MIGRATE_FAIL)
2124 p->numa_faults_locality[2] += pages;
2126 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2127 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2128 p->numa_faults_locality[local] += pages;
2131 static void reset_ptenuma_scan(struct task_struct *p)
2134 * We only did a read acquisition of the mmap sem, so
2135 * p->mm->numa_scan_seq is written to without exclusive access
2136 * and the update is not guaranteed to be atomic. That's not
2137 * much of an issue though, since this is just used for
2138 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2139 * expensive, to avoid any form of compiler optimizations:
2141 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2142 p->mm->numa_scan_offset = 0;
2146 * The expensive part of numa migration is done from task_work context.
2147 * Triggered from task_tick_numa().
2149 void task_numa_work(struct callback_head *work)
2151 unsigned long migrate, next_scan, now = jiffies;
2152 struct task_struct *p = current;
2153 struct mm_struct *mm = p->mm;
2154 struct vm_area_struct *vma;
2155 unsigned long start, end;
2156 unsigned long nr_pte_updates = 0;
2159 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2161 work->next = work; /* protect against double add */
2163 * Who cares about NUMA placement when they're dying.
2165 * NOTE: make sure not to dereference p->mm before this check,
2166 * exit_task_work() happens _after_ exit_mm() so we could be called
2167 * without p->mm even though we still had it when we enqueued this
2170 if (p->flags & PF_EXITING)
2173 if (!mm->numa_next_scan) {
2174 mm->numa_next_scan = now +
2175 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2179 * Enforce maximal scan/migration frequency..
2181 migrate = mm->numa_next_scan;
2182 if (time_before(now, migrate))
2185 if (p->numa_scan_period == 0) {
2186 p->numa_scan_period_max = task_scan_max(p);
2187 p->numa_scan_period = task_scan_min(p);
2190 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2191 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2195 * Delay this task enough that another task of this mm will likely win
2196 * the next time around.
2198 p->node_stamp += 2 * TICK_NSEC;
2200 start = mm->numa_scan_offset;
2201 pages = sysctl_numa_balancing_scan_size;
2202 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2206 down_read(&mm->mmap_sem);
2207 vma = find_vma(mm, start);
2209 reset_ptenuma_scan(p);
2213 for (; vma; vma = vma->vm_next) {
2214 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2215 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2220 * Shared library pages mapped by multiple processes are not
2221 * migrated as it is expected they are cache replicated. Avoid
2222 * hinting faults in read-only file-backed mappings or the vdso
2223 * as migrating the pages will be of marginal benefit.
2226 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2230 * Skip inaccessible VMAs to avoid any confusion between
2231 * PROT_NONE and NUMA hinting ptes
2233 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2237 start = max(start, vma->vm_start);
2238 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2239 end = min(end, vma->vm_end);
2240 nr_pte_updates += change_prot_numa(vma, start, end);
2243 * Scan sysctl_numa_balancing_scan_size but ensure that
2244 * at least one PTE is updated so that unused virtual
2245 * address space is quickly skipped.
2248 pages -= (end - start) >> PAGE_SHIFT;
2255 } while (end != vma->vm_end);
2260 * It is possible to reach the end of the VMA list but the last few
2261 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2262 * would find the !migratable VMA on the next scan but not reset the
2263 * scanner to the start so check it now.
2266 mm->numa_scan_offset = start;
2268 reset_ptenuma_scan(p);
2269 up_read(&mm->mmap_sem);
2273 * Drive the periodic memory faults..
2275 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2277 struct callback_head *work = &curr->numa_work;
2281 * We don't care about NUMA placement if we don't have memory.
2283 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2287 * Using runtime rather than walltime has the dual advantage that
2288 * we (mostly) drive the selection from busy threads and that the
2289 * task needs to have done some actual work before we bother with
2292 now = curr->se.sum_exec_runtime;
2293 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2295 if (now - curr->node_stamp > period) {
2296 if (!curr->node_stamp)
2297 curr->numa_scan_period = task_scan_min(curr);
2298 curr->node_stamp += period;
2300 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2301 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2302 task_work_add(curr, work, true);
2307 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2311 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2315 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2318 #endif /* CONFIG_NUMA_BALANCING */
2321 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2323 update_load_add(&cfs_rq->load, se->load.weight);
2324 if (!parent_entity(se))
2325 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2327 if (entity_is_task(se)) {
2328 struct rq *rq = rq_of(cfs_rq);
2330 account_numa_enqueue(rq, task_of(se));
2331 list_add(&se->group_node, &rq->cfs_tasks);
2334 cfs_rq->nr_running++;
2338 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2340 update_load_sub(&cfs_rq->load, se->load.weight);
2341 if (!parent_entity(se))
2342 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2343 if (entity_is_task(se)) {
2344 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2345 list_del_init(&se->group_node);
2347 cfs_rq->nr_running--;
2350 #ifdef CONFIG_FAIR_GROUP_SCHED
2352 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
2357 * Use this CPU's real-time load instead of the last load contribution
2358 * as the updating of the contribution is delayed, and we will use the
2359 * the real-time load to calc the share. See update_tg_load_avg().
2361 tg_weight = atomic_long_read(&tg->load_avg);
2362 tg_weight -= cfs_rq->tg_load_avg_contrib;
2363 tg_weight += cfs_rq->avg.load_avg;
2368 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2370 long tg_weight, load, shares;
2372 tg_weight = calc_tg_weight(tg, cfs_rq);
2373 load = cfs_rq->avg.load_avg;
2375 shares = (tg->shares * load);
2377 shares /= tg_weight;
2379 if (shares < MIN_SHARES)
2380 shares = MIN_SHARES;
2381 if (shares > tg->shares)
2382 shares = tg->shares;
2386 # else /* CONFIG_SMP */
2387 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2391 # endif /* CONFIG_SMP */
2392 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2393 unsigned long weight)
2396 /* commit outstanding execution time */
2397 if (cfs_rq->curr == se)
2398 update_curr(cfs_rq);
2399 account_entity_dequeue(cfs_rq, se);
2402 update_load_set(&se->load, weight);
2405 account_entity_enqueue(cfs_rq, se);
2408 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2410 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2412 struct task_group *tg;
2413 struct sched_entity *se;
2417 se = tg->se[cpu_of(rq_of(cfs_rq))];
2418 if (!se || throttled_hierarchy(cfs_rq))
2421 if (likely(se->load.weight == tg->shares))
2424 shares = calc_cfs_shares(cfs_rq, tg);
2426 reweight_entity(cfs_rq_of(se), se, shares);
2428 #else /* CONFIG_FAIR_GROUP_SCHED */
2429 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2432 #endif /* CONFIG_FAIR_GROUP_SCHED */
2435 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2436 static const u32 runnable_avg_yN_inv[] = {
2437 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2438 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2439 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2440 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2441 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2442 0x85aac367, 0x82cd8698,
2446 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
2447 * over-estimates when re-combining.
2449 static const u32 runnable_avg_yN_sum[] = {
2450 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2451 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2452 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2457 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
2459 static __always_inline u64 decay_load(u64 val, u64 n)
2461 unsigned int local_n;
2465 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2468 /* after bounds checking we can collapse to 32-bit */
2472 * As y^PERIOD = 1/2, we can combine
2473 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2474 * With a look-up table which covers y^n (n<PERIOD)
2476 * To achieve constant time decay_load.
2478 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2479 val >>= local_n / LOAD_AVG_PERIOD;
2480 local_n %= LOAD_AVG_PERIOD;
2483 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2488 * For updates fully spanning n periods, the contribution to runnable
2489 * average will be: \Sum 1024*y^n
2491 * We can compute this reasonably efficiently by combining:
2492 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2494 static u32 __compute_runnable_contrib(u64 n)
2498 if (likely(n <= LOAD_AVG_PERIOD))
2499 return runnable_avg_yN_sum[n];
2500 else if (unlikely(n >= LOAD_AVG_MAX_N))
2501 return LOAD_AVG_MAX;
2503 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2505 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2506 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2508 n -= LOAD_AVG_PERIOD;
2509 } while (n > LOAD_AVG_PERIOD);
2511 contrib = decay_load(contrib, n);
2512 return contrib + runnable_avg_yN_sum[n];
2516 * We can represent the historical contribution to runnable average as the
2517 * coefficients of a geometric series. To do this we sub-divide our runnable
2518 * history into segments of approximately 1ms (1024us); label the segment that
2519 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2521 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2523 * (now) (~1ms ago) (~2ms ago)
2525 * Let u_i denote the fraction of p_i that the entity was runnable.
2527 * We then designate the fractions u_i as our co-efficients, yielding the
2528 * following representation of historical load:
2529 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2531 * We choose y based on the with of a reasonably scheduling period, fixing:
2534 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2535 * approximately half as much as the contribution to load within the last ms
2538 * When a period "rolls over" and we have new u_0`, multiplying the previous
2539 * sum again by y is sufficient to update:
2540 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2541 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2543 static __always_inline int
2544 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2545 unsigned long weight, int running)
2549 int delta_w, decayed = 0;
2550 unsigned long scale_freq = arch_scale_freq_capacity(NULL, cpu);
2552 delta = now - sa->last_update_time;
2554 * This should only happen when time goes backwards, which it
2555 * unfortunately does during sched clock init when we swap over to TSC.
2557 if ((s64)delta < 0) {
2558 sa->last_update_time = now;
2563 * Use 1024ns as the unit of measurement since it's a reasonable
2564 * approximation of 1us and fast to compute.
2569 sa->last_update_time = now;
2571 /* delta_w is the amount already accumulated against our next period */
2572 delta_w = sa->period_contrib;
2573 if (delta + delta_w >= 1024) {
2576 /* how much left for next period will start over, we don't know yet */
2577 sa->period_contrib = 0;
2580 * Now that we know we're crossing a period boundary, figure
2581 * out how much from delta we need to complete the current
2582 * period and accrue it.
2584 delta_w = 1024 - delta_w;
2586 sa->load_sum += weight * delta_w;
2588 sa->util_sum += delta_w * scale_freq >> SCHED_CAPACITY_SHIFT;
2592 /* Figure out how many additional periods this update spans */
2593 periods = delta / 1024;
2596 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2597 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2599 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2600 contrib = __compute_runnable_contrib(periods);
2602 sa->load_sum += weight * contrib;
2604 sa->util_sum += contrib * scale_freq >> SCHED_CAPACITY_SHIFT;
2607 /* Remainder of delta accrued against u_0` */
2609 sa->load_sum += weight * delta;
2611 sa->util_sum += delta * scale_freq >> SCHED_CAPACITY_SHIFT;
2613 sa->period_contrib += delta;
2616 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2617 sa->util_avg = (sa->util_sum << SCHED_LOAD_SHIFT) / LOAD_AVG_MAX;
2623 #ifdef CONFIG_FAIR_GROUP_SCHED
2625 * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2626 * and effective_load (which is not done because it is too costly).
2628 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2630 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2632 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2633 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2634 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2638 #else /* CONFIG_FAIR_GROUP_SCHED */
2639 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2640 #endif /* CONFIG_FAIR_GROUP_SCHED */
2642 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2644 /* Group cfs_rq's load_avg is used for task_h_load and update_cfs_share */
2645 static inline int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
2648 struct sched_avg *sa = &cfs_rq->avg;
2650 if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2651 long r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2652 sa->load_avg = max_t(long, sa->load_avg - r, 0);
2653 sa->load_sum = max_t(s64, sa->load_sum - r * LOAD_AVG_MAX, 0);
2656 if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2657 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2658 sa->util_avg = max_t(long, sa->util_avg - r, 0);
2659 sa->util_sum = max_t(s32, sa->util_sum -
2660 ((r * LOAD_AVG_MAX) >> SCHED_LOAD_SHIFT), 0);
2663 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2664 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL);
2666 #ifndef CONFIG_64BIT
2668 cfs_rq->load_last_update_time_copy = sa->last_update_time;
2674 /* Update task and its cfs_rq load average */
2675 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2677 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2678 int cpu = cpu_of(rq_of(cfs_rq));
2679 u64 now = cfs_rq_clock_task(cfs_rq);
2682 * Track task load average for carrying it to new CPU after migrated, and
2683 * track group sched_entity load average for task_h_load calc in migration
2685 __update_load_avg(now, cpu, &se->avg,
2686 se->on_rq * scale_load_down(se->load.weight), cfs_rq->curr == se);
2688 if (update_cfs_rq_load_avg(now, cfs_rq) && update_tg)
2689 update_tg_load_avg(cfs_rq, 0);
2692 /* Add the load generated by se into cfs_rq's load average */
2694 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2696 struct sched_avg *sa = &se->avg;
2697 u64 now = cfs_rq_clock_task(cfs_rq);
2698 int migrated = 0, decayed;
2700 if (sa->last_update_time == 0) {
2701 sa->last_update_time = now;
2705 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2706 se->on_rq * scale_load_down(se->load.weight), cfs_rq->curr == se);
2709 decayed = update_cfs_rq_load_avg(now, cfs_rq);
2712 cfs_rq->avg.load_avg += sa->load_avg;
2713 cfs_rq->avg.load_sum += sa->load_sum;
2714 cfs_rq->avg.util_avg += sa->util_avg;
2715 cfs_rq->avg.util_sum += sa->util_sum;
2718 if (decayed || migrated)
2719 update_tg_load_avg(cfs_rq, 0);
2723 * Task first catches up with cfs_rq, and then subtract
2724 * itself from the cfs_rq (task must be off the queue now).
2726 void remove_entity_load_avg(struct sched_entity *se)
2728 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2729 u64 last_update_time;
2731 #ifndef CONFIG_64BIT
2732 u64 last_update_time_copy;
2735 last_update_time_copy = cfs_rq->load_last_update_time_copy;
2737 last_update_time = cfs_rq->avg.last_update_time;
2738 } while (last_update_time != last_update_time_copy);
2740 last_update_time = cfs_rq->avg.last_update_time;
2743 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0);
2744 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
2745 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
2749 * Update the rq's load with the elapsed running time before entering
2750 * idle. if the last scheduled task is not a CFS task, idle_enter will
2751 * be the only way to update the runnable statistic.
2753 void idle_enter_fair(struct rq *this_rq)
2758 * Update the rq's load with the elapsed idle time before a task is
2759 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2760 * be the only way to update the runnable statistic.
2762 void idle_exit_fair(struct rq *this_rq)
2766 static int idle_balance(struct rq *this_rq);
2768 #else /* CONFIG_SMP */
2770 static inline void update_load_avg(struct sched_entity *se, int update_tg) {}
2772 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
2773 static inline void remove_entity_load_avg(struct sched_entity *se) {}
2775 static inline int idle_balance(struct rq *rq)
2780 #endif /* CONFIG_SMP */
2782 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2784 #ifdef CONFIG_SCHEDSTATS
2785 struct task_struct *tsk = NULL;
2787 if (entity_is_task(se))
2790 if (se->statistics.sleep_start) {
2791 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2796 if (unlikely(delta > se->statistics.sleep_max))
2797 se->statistics.sleep_max = delta;
2799 se->statistics.sleep_start = 0;
2800 se->statistics.sum_sleep_runtime += delta;
2803 account_scheduler_latency(tsk, delta >> 10, 1);
2804 trace_sched_stat_sleep(tsk, delta);
2807 if (se->statistics.block_start) {
2808 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2813 if (unlikely(delta > se->statistics.block_max))
2814 se->statistics.block_max = delta;
2816 se->statistics.block_start = 0;
2817 se->statistics.sum_sleep_runtime += delta;
2820 if (tsk->in_iowait) {
2821 se->statistics.iowait_sum += delta;
2822 se->statistics.iowait_count++;
2823 trace_sched_stat_iowait(tsk, delta);
2826 trace_sched_stat_blocked(tsk, delta);
2829 * Blocking time is in units of nanosecs, so shift by
2830 * 20 to get a milliseconds-range estimation of the
2831 * amount of time that the task spent sleeping:
2833 if (unlikely(prof_on == SLEEP_PROFILING)) {
2834 profile_hits(SLEEP_PROFILING,
2835 (void *)get_wchan(tsk),
2838 account_scheduler_latency(tsk, delta >> 10, 0);
2844 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2846 #ifdef CONFIG_SCHED_DEBUG
2847 s64 d = se->vruntime - cfs_rq->min_vruntime;
2852 if (d > 3*sysctl_sched_latency)
2853 schedstat_inc(cfs_rq, nr_spread_over);
2858 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2860 u64 vruntime = cfs_rq->min_vruntime;
2863 * The 'current' period is already promised to the current tasks,
2864 * however the extra weight of the new task will slow them down a
2865 * little, place the new task so that it fits in the slot that
2866 * stays open at the end.
2868 if (initial && sched_feat(START_DEBIT))
2869 vruntime += sched_vslice(cfs_rq, se);
2871 /* sleeps up to a single latency don't count. */
2873 unsigned long thresh = sysctl_sched_latency;
2876 * Halve their sleep time's effect, to allow
2877 * for a gentler effect of sleepers:
2879 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2885 /* ensure we never gain time by being placed backwards. */
2886 se->vruntime = max_vruntime(se->vruntime, vruntime);
2889 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2892 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2895 * Update the normalized vruntime before updating min_vruntime
2896 * through calling update_curr().
2898 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2899 se->vruntime += cfs_rq->min_vruntime;
2902 * Update run-time statistics of the 'current'.
2904 update_curr(cfs_rq);
2905 enqueue_entity_load_avg(cfs_rq, se);
2906 account_entity_enqueue(cfs_rq, se);
2907 update_cfs_shares(cfs_rq);
2909 if (flags & ENQUEUE_WAKEUP) {
2910 place_entity(cfs_rq, se, 0);
2911 enqueue_sleeper(cfs_rq, se);
2914 update_stats_enqueue(cfs_rq, se);
2915 check_spread(cfs_rq, se);
2916 if (se != cfs_rq->curr)
2917 __enqueue_entity(cfs_rq, se);
2920 if (cfs_rq->nr_running == 1) {
2921 list_add_leaf_cfs_rq(cfs_rq);
2922 check_enqueue_throttle(cfs_rq);
2926 static void __clear_buddies_last(struct sched_entity *se)
2928 for_each_sched_entity(se) {
2929 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2930 if (cfs_rq->last != se)
2933 cfs_rq->last = NULL;
2937 static void __clear_buddies_next(struct sched_entity *se)
2939 for_each_sched_entity(se) {
2940 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2941 if (cfs_rq->next != se)
2944 cfs_rq->next = NULL;
2948 static void __clear_buddies_skip(struct sched_entity *se)
2950 for_each_sched_entity(se) {
2951 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2952 if (cfs_rq->skip != se)
2955 cfs_rq->skip = NULL;
2959 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2961 if (cfs_rq->last == se)
2962 __clear_buddies_last(se);
2964 if (cfs_rq->next == se)
2965 __clear_buddies_next(se);
2967 if (cfs_rq->skip == se)
2968 __clear_buddies_skip(se);
2971 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2974 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2977 * Update run-time statistics of the 'current'.
2979 update_curr(cfs_rq);
2980 update_load_avg(se, 1);
2982 update_stats_dequeue(cfs_rq, se);
2983 if (flags & DEQUEUE_SLEEP) {
2984 #ifdef CONFIG_SCHEDSTATS
2985 if (entity_is_task(se)) {
2986 struct task_struct *tsk = task_of(se);
2988 if (tsk->state & TASK_INTERRUPTIBLE)
2989 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2990 if (tsk->state & TASK_UNINTERRUPTIBLE)
2991 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2996 clear_buddies(cfs_rq, se);
2998 if (se != cfs_rq->curr)
2999 __dequeue_entity(cfs_rq, se);
3001 account_entity_dequeue(cfs_rq, se);
3004 * Normalize the entity after updating the min_vruntime because the
3005 * update can refer to the ->curr item and we need to reflect this
3006 * movement in our normalized position.
3008 if (!(flags & DEQUEUE_SLEEP))
3009 se->vruntime -= cfs_rq->min_vruntime;
3011 /* return excess runtime on last dequeue */
3012 return_cfs_rq_runtime(cfs_rq);
3014 update_min_vruntime(cfs_rq);
3015 update_cfs_shares(cfs_rq);
3019 * Preempt the current task with a newly woken task if needed:
3022 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3024 unsigned long ideal_runtime, delta_exec;
3025 struct sched_entity *se;
3028 ideal_runtime = sched_slice(cfs_rq, curr);
3029 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3030 if (delta_exec > ideal_runtime) {
3031 resched_curr(rq_of(cfs_rq));
3033 * The current task ran long enough, ensure it doesn't get
3034 * re-elected due to buddy favours.
3036 clear_buddies(cfs_rq, curr);
3041 * Ensure that a task that missed wakeup preemption by a
3042 * narrow margin doesn't have to wait for a full slice.
3043 * This also mitigates buddy induced latencies under load.
3045 if (delta_exec < sysctl_sched_min_granularity)
3048 se = __pick_first_entity(cfs_rq);
3049 delta = curr->vruntime - se->vruntime;
3054 if (delta > ideal_runtime)
3055 resched_curr(rq_of(cfs_rq));
3059 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3061 /* 'current' is not kept within the tree. */
3064 * Any task has to be enqueued before it get to execute on
3065 * a CPU. So account for the time it spent waiting on the
3068 update_stats_wait_end(cfs_rq, se);
3069 __dequeue_entity(cfs_rq, se);
3070 update_load_avg(se, 1);
3073 update_stats_curr_start(cfs_rq, se);
3075 #ifdef CONFIG_SCHEDSTATS
3077 * Track our maximum slice length, if the CPU's load is at
3078 * least twice that of our own weight (i.e. dont track it
3079 * when there are only lesser-weight tasks around):
3081 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3082 se->statistics.slice_max = max(se->statistics.slice_max,
3083 se->sum_exec_runtime - se->prev_sum_exec_runtime);
3086 se->prev_sum_exec_runtime = se->sum_exec_runtime;
3090 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3093 * Pick the next process, keeping these things in mind, in this order:
3094 * 1) keep things fair between processes/task groups
3095 * 2) pick the "next" process, since someone really wants that to run
3096 * 3) pick the "last" process, for cache locality
3097 * 4) do not run the "skip" process, if something else is available
3099 static struct sched_entity *
3100 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3102 struct sched_entity *left = __pick_first_entity(cfs_rq);
3103 struct sched_entity *se;
3106 * If curr is set we have to see if its left of the leftmost entity
3107 * still in the tree, provided there was anything in the tree at all.
3109 if (!left || (curr && entity_before(curr, left)))
3112 se = left; /* ideally we run the leftmost entity */
3115 * Avoid running the skip buddy, if running something else can
3116 * be done without getting too unfair.
3118 if (cfs_rq->skip == se) {
3119 struct sched_entity *second;
3122 second = __pick_first_entity(cfs_rq);
3124 second = __pick_next_entity(se);
3125 if (!second || (curr && entity_before(curr, second)))
3129 if (second && wakeup_preempt_entity(second, left) < 1)
3134 * Prefer last buddy, try to return the CPU to a preempted task.
3136 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3140 * Someone really wants this to run. If it's not unfair, run it.
3142 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3145 clear_buddies(cfs_rq, se);
3150 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3152 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3155 * If still on the runqueue then deactivate_task()
3156 * was not called and update_curr() has to be done:
3159 update_curr(cfs_rq);
3161 /* throttle cfs_rqs exceeding runtime */
3162 check_cfs_rq_runtime(cfs_rq);
3164 check_spread(cfs_rq, prev);
3166 update_stats_wait_start(cfs_rq, prev);
3167 /* Put 'current' back into the tree. */
3168 __enqueue_entity(cfs_rq, prev);
3169 /* in !on_rq case, update occurred at dequeue */
3170 update_load_avg(prev, 0);
3172 cfs_rq->curr = NULL;
3176 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3179 * Update run-time statistics of the 'current'.
3181 update_curr(cfs_rq);
3184 * Ensure that runnable average is periodically updated.
3186 update_load_avg(curr, 1);
3187 update_cfs_shares(cfs_rq);
3189 #ifdef CONFIG_SCHED_HRTICK
3191 * queued ticks are scheduled to match the slice, so don't bother
3192 * validating it and just reschedule.
3195 resched_curr(rq_of(cfs_rq));
3199 * don't let the period tick interfere with the hrtick preemption
3201 if (!sched_feat(DOUBLE_TICK) &&
3202 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3206 if (cfs_rq->nr_running > 1)
3207 check_preempt_tick(cfs_rq, curr);
3211 /**************************************************
3212 * CFS bandwidth control machinery
3215 #ifdef CONFIG_CFS_BANDWIDTH
3217 #ifdef HAVE_JUMP_LABEL
3218 static struct static_key __cfs_bandwidth_used;
3220 static inline bool cfs_bandwidth_used(void)
3222 return static_key_false(&__cfs_bandwidth_used);
3225 void cfs_bandwidth_usage_inc(void)
3227 static_key_slow_inc(&__cfs_bandwidth_used);
3230 void cfs_bandwidth_usage_dec(void)
3232 static_key_slow_dec(&__cfs_bandwidth_used);
3234 #else /* HAVE_JUMP_LABEL */
3235 static bool cfs_bandwidth_used(void)
3240 void cfs_bandwidth_usage_inc(void) {}
3241 void cfs_bandwidth_usage_dec(void) {}
3242 #endif /* HAVE_JUMP_LABEL */
3245 * default period for cfs group bandwidth.
3246 * default: 0.1s, units: nanoseconds
3248 static inline u64 default_cfs_period(void)
3250 return 100000000ULL;
3253 static inline u64 sched_cfs_bandwidth_slice(void)
3255 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3259 * Replenish runtime according to assigned quota and update expiration time.
3260 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3261 * additional synchronization around rq->lock.
3263 * requires cfs_b->lock
3265 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3269 if (cfs_b->quota == RUNTIME_INF)
3272 now = sched_clock_cpu(smp_processor_id());
3273 cfs_b->runtime = cfs_b->quota;
3274 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3277 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3279 return &tg->cfs_bandwidth;
3282 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3283 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3285 if (unlikely(cfs_rq->throttle_count))
3286 return cfs_rq->throttled_clock_task;
3288 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3291 /* returns 0 on failure to allocate runtime */
3292 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3294 struct task_group *tg = cfs_rq->tg;
3295 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3296 u64 amount = 0, min_amount, expires;
3298 /* note: this is a positive sum as runtime_remaining <= 0 */
3299 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3301 raw_spin_lock(&cfs_b->lock);
3302 if (cfs_b->quota == RUNTIME_INF)
3303 amount = min_amount;
3305 start_cfs_bandwidth(cfs_b);
3307 if (cfs_b->runtime > 0) {
3308 amount = min(cfs_b->runtime, min_amount);
3309 cfs_b->runtime -= amount;
3313 expires = cfs_b->runtime_expires;
3314 raw_spin_unlock(&cfs_b->lock);
3316 cfs_rq->runtime_remaining += amount;
3318 * we may have advanced our local expiration to account for allowed
3319 * spread between our sched_clock and the one on which runtime was
3322 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3323 cfs_rq->runtime_expires = expires;
3325 return cfs_rq->runtime_remaining > 0;
3329 * Note: This depends on the synchronization provided by sched_clock and the
3330 * fact that rq->clock snapshots this value.
3332 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3334 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3336 /* if the deadline is ahead of our clock, nothing to do */
3337 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3340 if (cfs_rq->runtime_remaining < 0)
3344 * If the local deadline has passed we have to consider the
3345 * possibility that our sched_clock is 'fast' and the global deadline
3346 * has not truly expired.
3348 * Fortunately we can check determine whether this the case by checking
3349 * whether the global deadline has advanced. It is valid to compare
3350 * cfs_b->runtime_expires without any locks since we only care about
3351 * exact equality, so a partial write will still work.
3354 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3355 /* extend local deadline, drift is bounded above by 2 ticks */
3356 cfs_rq->runtime_expires += TICK_NSEC;
3358 /* global deadline is ahead, expiration has passed */
3359 cfs_rq->runtime_remaining = 0;
3363 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3365 /* dock delta_exec before expiring quota (as it could span periods) */
3366 cfs_rq->runtime_remaining -= delta_exec;
3367 expire_cfs_rq_runtime(cfs_rq);
3369 if (likely(cfs_rq->runtime_remaining > 0))
3373 * if we're unable to extend our runtime we resched so that the active
3374 * hierarchy can be throttled
3376 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3377 resched_curr(rq_of(cfs_rq));
3380 static __always_inline
3381 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3383 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3386 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3389 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3391 return cfs_bandwidth_used() && cfs_rq->throttled;
3394 /* check whether cfs_rq, or any parent, is throttled */
3395 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3397 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3401 * Ensure that neither of the group entities corresponding to src_cpu or
3402 * dest_cpu are members of a throttled hierarchy when performing group
3403 * load-balance operations.
3405 static inline int throttled_lb_pair(struct task_group *tg,
3406 int src_cpu, int dest_cpu)
3408 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3410 src_cfs_rq = tg->cfs_rq[src_cpu];
3411 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3413 return throttled_hierarchy(src_cfs_rq) ||
3414 throttled_hierarchy(dest_cfs_rq);
3417 /* updated child weight may affect parent so we have to do this bottom up */
3418 static int tg_unthrottle_up(struct task_group *tg, void *data)
3420 struct rq *rq = data;
3421 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3423 cfs_rq->throttle_count--;
3425 if (!cfs_rq->throttle_count) {
3426 /* adjust cfs_rq_clock_task() */
3427 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3428 cfs_rq->throttled_clock_task;
3435 static int tg_throttle_down(struct task_group *tg, void *data)
3437 struct rq *rq = data;
3438 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3440 /* group is entering throttled state, stop time */
3441 if (!cfs_rq->throttle_count)
3442 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3443 cfs_rq->throttle_count++;
3448 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3450 struct rq *rq = rq_of(cfs_rq);
3451 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3452 struct sched_entity *se;
3453 long task_delta, dequeue = 1;
3456 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3458 /* freeze hierarchy runnable averages while throttled */
3460 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3463 task_delta = cfs_rq->h_nr_running;
3464 for_each_sched_entity(se) {
3465 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3466 /* throttled entity or throttle-on-deactivate */
3471 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3472 qcfs_rq->h_nr_running -= task_delta;
3474 if (qcfs_rq->load.weight)
3479 sub_nr_running(rq, task_delta);
3481 cfs_rq->throttled = 1;
3482 cfs_rq->throttled_clock = rq_clock(rq);
3483 raw_spin_lock(&cfs_b->lock);
3484 empty = list_empty(&cfs_b->throttled_cfs_rq);
3487 * Add to the _head_ of the list, so that an already-started
3488 * distribute_cfs_runtime will not see us
3490 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3493 * If we're the first throttled task, make sure the bandwidth
3497 start_cfs_bandwidth(cfs_b);
3499 raw_spin_unlock(&cfs_b->lock);
3502 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3504 struct rq *rq = rq_of(cfs_rq);
3505 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3506 struct sched_entity *se;
3510 se = cfs_rq->tg->se[cpu_of(rq)];
3512 cfs_rq->throttled = 0;
3514 update_rq_clock(rq);
3516 raw_spin_lock(&cfs_b->lock);
3517 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3518 list_del_rcu(&cfs_rq->throttled_list);
3519 raw_spin_unlock(&cfs_b->lock);
3521 /* update hierarchical throttle state */
3522 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3524 if (!cfs_rq->load.weight)
3527 task_delta = cfs_rq->h_nr_running;
3528 for_each_sched_entity(se) {
3532 cfs_rq = cfs_rq_of(se);
3534 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3535 cfs_rq->h_nr_running += task_delta;
3537 if (cfs_rq_throttled(cfs_rq))
3542 add_nr_running(rq, task_delta);
3544 /* determine whether we need to wake up potentially idle cpu */
3545 if (rq->curr == rq->idle && rq->cfs.nr_running)
3549 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3550 u64 remaining, u64 expires)
3552 struct cfs_rq *cfs_rq;
3554 u64 starting_runtime = remaining;
3557 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3559 struct rq *rq = rq_of(cfs_rq);
3561 raw_spin_lock(&rq->lock);
3562 if (!cfs_rq_throttled(cfs_rq))
3565 runtime = -cfs_rq->runtime_remaining + 1;
3566 if (runtime > remaining)
3567 runtime = remaining;
3568 remaining -= runtime;
3570 cfs_rq->runtime_remaining += runtime;
3571 cfs_rq->runtime_expires = expires;
3573 /* we check whether we're throttled above */
3574 if (cfs_rq->runtime_remaining > 0)
3575 unthrottle_cfs_rq(cfs_rq);
3578 raw_spin_unlock(&rq->lock);
3585 return starting_runtime - remaining;
3589 * Responsible for refilling a task_group's bandwidth and unthrottling its
3590 * cfs_rqs as appropriate. If there has been no activity within the last
3591 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3592 * used to track this state.
3594 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3596 u64 runtime, runtime_expires;
3599 /* no need to continue the timer with no bandwidth constraint */
3600 if (cfs_b->quota == RUNTIME_INF)
3601 goto out_deactivate;
3603 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3604 cfs_b->nr_periods += overrun;
3607 * idle depends on !throttled (for the case of a large deficit), and if
3608 * we're going inactive then everything else can be deferred
3610 if (cfs_b->idle && !throttled)
3611 goto out_deactivate;
3613 __refill_cfs_bandwidth_runtime(cfs_b);
3616 /* mark as potentially idle for the upcoming period */
3621 /* account preceding periods in which throttling occurred */
3622 cfs_b->nr_throttled += overrun;
3624 runtime_expires = cfs_b->runtime_expires;
3627 * This check is repeated as we are holding onto the new bandwidth while
3628 * we unthrottle. This can potentially race with an unthrottled group
3629 * trying to acquire new bandwidth from the global pool. This can result
3630 * in us over-using our runtime if it is all used during this loop, but
3631 * only by limited amounts in that extreme case.
3633 while (throttled && cfs_b->runtime > 0) {
3634 runtime = cfs_b->runtime;
3635 raw_spin_unlock(&cfs_b->lock);
3636 /* we can't nest cfs_b->lock while distributing bandwidth */
3637 runtime = distribute_cfs_runtime(cfs_b, runtime,
3639 raw_spin_lock(&cfs_b->lock);
3641 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3643 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3647 * While we are ensured activity in the period following an
3648 * unthrottle, this also covers the case in which the new bandwidth is
3649 * insufficient to cover the existing bandwidth deficit. (Forcing the
3650 * timer to remain active while there are any throttled entities.)
3660 /* a cfs_rq won't donate quota below this amount */
3661 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3662 /* minimum remaining period time to redistribute slack quota */
3663 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3664 /* how long we wait to gather additional slack before distributing */
3665 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3668 * Are we near the end of the current quota period?
3670 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3671 * hrtimer base being cleared by hrtimer_start. In the case of
3672 * migrate_hrtimers, base is never cleared, so we are fine.
3674 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3676 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3679 /* if the call-back is running a quota refresh is already occurring */
3680 if (hrtimer_callback_running(refresh_timer))
3683 /* is a quota refresh about to occur? */
3684 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3685 if (remaining < min_expire)
3691 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3693 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3695 /* if there's a quota refresh soon don't bother with slack */
3696 if (runtime_refresh_within(cfs_b, min_left))
3699 hrtimer_start(&cfs_b->slack_timer,
3700 ns_to_ktime(cfs_bandwidth_slack_period),
3704 /* we know any runtime found here is valid as update_curr() precedes return */
3705 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3707 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3708 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3710 if (slack_runtime <= 0)
3713 raw_spin_lock(&cfs_b->lock);
3714 if (cfs_b->quota != RUNTIME_INF &&
3715 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3716 cfs_b->runtime += slack_runtime;
3718 /* we are under rq->lock, defer unthrottling using a timer */
3719 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3720 !list_empty(&cfs_b->throttled_cfs_rq))
3721 start_cfs_slack_bandwidth(cfs_b);
3723 raw_spin_unlock(&cfs_b->lock);
3725 /* even if it's not valid for return we don't want to try again */
3726 cfs_rq->runtime_remaining -= slack_runtime;
3729 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3731 if (!cfs_bandwidth_used())
3734 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3737 __return_cfs_rq_runtime(cfs_rq);
3741 * This is done with a timer (instead of inline with bandwidth return) since
3742 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3744 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3746 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3749 /* confirm we're still not at a refresh boundary */
3750 raw_spin_lock(&cfs_b->lock);
3751 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3752 raw_spin_unlock(&cfs_b->lock);
3756 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
3757 runtime = cfs_b->runtime;
3759 expires = cfs_b->runtime_expires;
3760 raw_spin_unlock(&cfs_b->lock);
3765 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3767 raw_spin_lock(&cfs_b->lock);
3768 if (expires == cfs_b->runtime_expires)
3769 cfs_b->runtime -= min(runtime, cfs_b->runtime);
3770 raw_spin_unlock(&cfs_b->lock);
3774 * When a group wakes up we want to make sure that its quota is not already
3775 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3776 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3778 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3780 if (!cfs_bandwidth_used())
3783 /* an active group must be handled by the update_curr()->put() path */
3784 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3787 /* ensure the group is not already throttled */
3788 if (cfs_rq_throttled(cfs_rq))
3791 /* update runtime allocation */
3792 account_cfs_rq_runtime(cfs_rq, 0);
3793 if (cfs_rq->runtime_remaining <= 0)
3794 throttle_cfs_rq(cfs_rq);
3797 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3798 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3800 if (!cfs_bandwidth_used())
3803 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3807 * it's possible for a throttled entity to be forced into a running
3808 * state (e.g. set_curr_task), in this case we're finished.
3810 if (cfs_rq_throttled(cfs_rq))
3813 throttle_cfs_rq(cfs_rq);
3817 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3819 struct cfs_bandwidth *cfs_b =
3820 container_of(timer, struct cfs_bandwidth, slack_timer);
3822 do_sched_cfs_slack_timer(cfs_b);
3824 return HRTIMER_NORESTART;
3827 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3829 struct cfs_bandwidth *cfs_b =
3830 container_of(timer, struct cfs_bandwidth, period_timer);
3834 raw_spin_lock(&cfs_b->lock);
3836 overrun = hrtimer_forward_now(timer, cfs_b->period);
3840 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3843 cfs_b->period_active = 0;
3844 raw_spin_unlock(&cfs_b->lock);
3846 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3849 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3851 raw_spin_lock_init(&cfs_b->lock);
3853 cfs_b->quota = RUNTIME_INF;
3854 cfs_b->period = ns_to_ktime(default_cfs_period());
3856 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3857 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
3858 cfs_b->period_timer.function = sched_cfs_period_timer;
3859 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3860 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3863 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3865 cfs_rq->runtime_enabled = 0;
3866 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3869 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3871 lockdep_assert_held(&cfs_b->lock);
3873 if (!cfs_b->period_active) {
3874 cfs_b->period_active = 1;
3875 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
3876 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
3880 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3882 /* init_cfs_bandwidth() was not called */
3883 if (!cfs_b->throttled_cfs_rq.next)
3886 hrtimer_cancel(&cfs_b->period_timer);
3887 hrtimer_cancel(&cfs_b->slack_timer);
3890 static void __maybe_unused update_runtime_enabled(struct rq *rq)
3892 struct cfs_rq *cfs_rq;
3894 for_each_leaf_cfs_rq(rq, cfs_rq) {
3895 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
3897 raw_spin_lock(&cfs_b->lock);
3898 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
3899 raw_spin_unlock(&cfs_b->lock);
3903 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3905 struct cfs_rq *cfs_rq;
3907 for_each_leaf_cfs_rq(rq, cfs_rq) {
3908 if (!cfs_rq->runtime_enabled)
3912 * clock_task is not advancing so we just need to make sure
3913 * there's some valid quota amount
3915 cfs_rq->runtime_remaining = 1;
3917 * Offline rq is schedulable till cpu is completely disabled
3918 * in take_cpu_down(), so we prevent new cfs throttling here.
3920 cfs_rq->runtime_enabled = 0;
3922 if (cfs_rq_throttled(cfs_rq))
3923 unthrottle_cfs_rq(cfs_rq);
3927 #else /* CONFIG_CFS_BANDWIDTH */
3928 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3930 return rq_clock_task(rq_of(cfs_rq));
3933 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
3934 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
3935 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3936 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3938 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3943 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3948 static inline int throttled_lb_pair(struct task_group *tg,
3949 int src_cpu, int dest_cpu)
3954 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3956 #ifdef CONFIG_FAIR_GROUP_SCHED
3957 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3960 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3964 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3965 static inline void update_runtime_enabled(struct rq *rq) {}
3966 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3968 #endif /* CONFIG_CFS_BANDWIDTH */
3970 /**************************************************
3971 * CFS operations on tasks:
3974 #ifdef CONFIG_SCHED_HRTICK
3975 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3977 struct sched_entity *se = &p->se;
3978 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3980 WARN_ON(task_rq(p) != rq);
3982 if (cfs_rq->nr_running > 1) {
3983 u64 slice = sched_slice(cfs_rq, se);
3984 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3985 s64 delta = slice - ran;
3992 hrtick_start(rq, delta);
3997 * called from enqueue/dequeue and updates the hrtick when the
3998 * current task is from our class and nr_running is low enough
4001 static void hrtick_update(struct rq *rq)
4003 struct task_struct *curr = rq->curr;
4005 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4008 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4009 hrtick_start_fair(rq, curr);
4011 #else /* !CONFIG_SCHED_HRTICK */
4013 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4017 static inline void hrtick_update(struct rq *rq)
4023 * The enqueue_task method is called before nr_running is
4024 * increased. Here we update the fair scheduling stats and
4025 * then put the task into the rbtree:
4028 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4030 struct cfs_rq *cfs_rq;
4031 struct sched_entity *se = &p->se;
4033 for_each_sched_entity(se) {
4036 cfs_rq = cfs_rq_of(se);
4037 enqueue_entity(cfs_rq, se, flags);
4040 * end evaluation on encountering a throttled cfs_rq
4042 * note: in the case of encountering a throttled cfs_rq we will
4043 * post the final h_nr_running increment below.
4045 if (cfs_rq_throttled(cfs_rq))
4047 cfs_rq->h_nr_running++;
4049 flags = ENQUEUE_WAKEUP;
4052 for_each_sched_entity(se) {
4053 cfs_rq = cfs_rq_of(se);
4054 cfs_rq->h_nr_running++;
4056 if (cfs_rq_throttled(cfs_rq))
4059 update_load_avg(se, 1);
4060 update_cfs_shares(cfs_rq);
4064 add_nr_running(rq, 1);
4069 static void set_next_buddy(struct sched_entity *se);
4072 * The dequeue_task method is called before nr_running is
4073 * decreased. We remove the task from the rbtree and
4074 * update the fair scheduling stats:
4076 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4078 struct cfs_rq *cfs_rq;
4079 struct sched_entity *se = &p->se;
4080 int task_sleep = flags & DEQUEUE_SLEEP;
4082 for_each_sched_entity(se) {
4083 cfs_rq = cfs_rq_of(se);
4084 dequeue_entity(cfs_rq, se, flags);
4087 * end evaluation on encountering a throttled cfs_rq
4089 * note: in the case of encountering a throttled cfs_rq we will
4090 * post the final h_nr_running decrement below.
4092 if (cfs_rq_throttled(cfs_rq))
4094 cfs_rq->h_nr_running--;
4096 /* Don't dequeue parent if it has other entities besides us */
4097 if (cfs_rq->load.weight) {
4099 * Bias pick_next to pick a task from this cfs_rq, as
4100 * p is sleeping when it is within its sched_slice.
4102 if (task_sleep && parent_entity(se))
4103 set_next_buddy(parent_entity(se));
4105 /* avoid re-evaluating load for this entity */
4106 se = parent_entity(se);
4109 flags |= DEQUEUE_SLEEP;
4112 for_each_sched_entity(se) {
4113 cfs_rq = cfs_rq_of(se);
4114 cfs_rq->h_nr_running--;
4116 if (cfs_rq_throttled(cfs_rq))
4119 update_load_avg(se, 1);
4120 update_cfs_shares(cfs_rq);
4124 sub_nr_running(rq, 1);
4132 * per rq 'load' arrray crap; XXX kill this.
4136 * The exact cpuload at various idx values, calculated at every tick would be
4137 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
4139 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
4140 * on nth tick when cpu may be busy, then we have:
4141 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4142 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
4144 * decay_load_missed() below does efficient calculation of
4145 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
4146 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
4148 * The calculation is approximated on a 128 point scale.
4149 * degrade_zero_ticks is the number of ticks after which load at any
4150 * particular idx is approximated to be zero.
4151 * degrade_factor is a precomputed table, a row for each load idx.
4152 * Each column corresponds to degradation factor for a power of two ticks,
4153 * based on 128 point scale.
4155 * row 2, col 3 (=12) says that the degradation at load idx 2 after
4156 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
4158 * With this power of 2 load factors, we can degrade the load n times
4159 * by looking at 1 bits in n and doing as many mult/shift instead of
4160 * n mult/shifts needed by the exact degradation.
4162 #define DEGRADE_SHIFT 7
4163 static const unsigned char
4164 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4165 static const unsigned char
4166 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4167 {0, 0, 0, 0, 0, 0, 0, 0},
4168 {64, 32, 8, 0, 0, 0, 0, 0},
4169 {96, 72, 40, 12, 1, 0, 0},
4170 {112, 98, 75, 43, 15, 1, 0},
4171 {120, 112, 98, 76, 45, 16, 2} };
4174 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4175 * would be when CPU is idle and so we just decay the old load without
4176 * adding any new load.
4178 static unsigned long
4179 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4183 if (!missed_updates)
4186 if (missed_updates >= degrade_zero_ticks[idx])
4190 return load >> missed_updates;
4192 while (missed_updates) {
4193 if (missed_updates % 2)
4194 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4196 missed_updates >>= 1;
4203 * Update rq->cpu_load[] statistics. This function is usually called every
4204 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
4205 * every tick. We fix it up based on jiffies.
4207 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
4208 unsigned long pending_updates)
4212 this_rq->nr_load_updates++;
4214 /* Update our load: */
4215 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4216 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4217 unsigned long old_load, new_load;
4219 /* scale is effectively 1 << i now, and >> i divides by scale */
4221 old_load = this_rq->cpu_load[i];
4222 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4223 new_load = this_load;
4225 * Round up the averaging division if load is increasing. This
4226 * prevents us from getting stuck on 9 if the load is 10, for
4229 if (new_load > old_load)
4230 new_load += scale - 1;
4232 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4235 sched_avg_update(this_rq);
4238 #ifdef CONFIG_NO_HZ_COMMON
4240 * There is no sane way to deal with nohz on smp when using jiffies because the
4241 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4242 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4244 * Therefore we cannot use the delta approach from the regular tick since that
4245 * would seriously skew the load calculation. However we'll make do for those
4246 * updates happening while idle (nohz_idle_balance) or coming out of idle
4247 * (tick_nohz_idle_exit).
4249 * This means we might still be one tick off for nohz periods.
4253 * Called from nohz_idle_balance() to update the load ratings before doing the
4256 static void update_idle_cpu_load(struct rq *this_rq)
4258 unsigned long curr_jiffies = READ_ONCE(jiffies);
4259 unsigned long load = this_rq->cfs.avg.load_avg;
4260 unsigned long pending_updates;
4263 * bail if there's load or we're actually up-to-date.
4265 if (load || curr_jiffies == this_rq->last_load_update_tick)
4268 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4269 this_rq->last_load_update_tick = curr_jiffies;
4271 __update_cpu_load(this_rq, load, pending_updates);
4275 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
4277 void update_cpu_load_nohz(void)
4279 struct rq *this_rq = this_rq();
4280 unsigned long curr_jiffies = READ_ONCE(jiffies);
4281 unsigned long pending_updates;
4283 if (curr_jiffies == this_rq->last_load_update_tick)
4286 raw_spin_lock(&this_rq->lock);
4287 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4288 if (pending_updates) {
4289 this_rq->last_load_update_tick = curr_jiffies;
4291 * We were idle, this means load 0, the current load might be
4292 * !0 due to remote wakeups and the sort.
4294 __update_cpu_load(this_rq, 0, pending_updates);
4296 raw_spin_unlock(&this_rq->lock);
4298 #endif /* CONFIG_NO_HZ */
4301 * Called from scheduler_tick()
4303 void update_cpu_load_active(struct rq *this_rq)
4305 unsigned long load = this_rq->cfs.avg.load_avg;
4307 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
4309 this_rq->last_load_update_tick = jiffies;
4310 __update_cpu_load(this_rq, load, 1);
4313 /* Used instead of source_load when we know the type == 0 */
4314 static unsigned long weighted_cpuload(const int cpu)
4316 return cpu_rq(cpu)->cfs.avg.load_avg;
4320 * Return a low guess at the load of a migration-source cpu weighted
4321 * according to the scheduling class and "nice" value.
4323 * We want to under-estimate the load of migration sources, to
4324 * balance conservatively.
4326 static unsigned long source_load(int cpu, int type)
4328 struct rq *rq = cpu_rq(cpu);
4329 unsigned long total = weighted_cpuload(cpu);
4331 if (type == 0 || !sched_feat(LB_BIAS))
4334 return min(rq->cpu_load[type-1], total);
4338 * Return a high guess at the load of a migration-target cpu weighted
4339 * according to the scheduling class and "nice" value.
4341 static unsigned long target_load(int cpu, int type)
4343 struct rq *rq = cpu_rq(cpu);
4344 unsigned long total = weighted_cpuload(cpu);
4346 if (type == 0 || !sched_feat(LB_BIAS))
4349 return max(rq->cpu_load[type-1], total);
4352 static unsigned long capacity_of(int cpu)
4354 return cpu_rq(cpu)->cpu_capacity;
4357 static unsigned long capacity_orig_of(int cpu)
4359 return cpu_rq(cpu)->cpu_capacity_orig;
4362 static unsigned long cpu_avg_load_per_task(int cpu)
4364 struct rq *rq = cpu_rq(cpu);
4365 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4366 unsigned long load_avg = rq->cfs.avg.load_avg;
4369 return load_avg / nr_running;
4374 static void record_wakee(struct task_struct *p)
4377 * Rough decay (wiping) for cost saving, don't worry
4378 * about the boundary, really active task won't care
4381 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
4382 current->wakee_flips >>= 1;
4383 current->wakee_flip_decay_ts = jiffies;
4386 if (current->last_wakee != p) {
4387 current->last_wakee = p;
4388 current->wakee_flips++;
4392 static void task_waking_fair(struct task_struct *p)
4394 struct sched_entity *se = &p->se;
4395 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4398 #ifndef CONFIG_64BIT
4399 u64 min_vruntime_copy;
4402 min_vruntime_copy = cfs_rq->min_vruntime_copy;
4404 min_vruntime = cfs_rq->min_vruntime;
4405 } while (min_vruntime != min_vruntime_copy);
4407 min_vruntime = cfs_rq->min_vruntime;
4410 se->vruntime -= min_vruntime;
4414 #ifdef CONFIG_FAIR_GROUP_SCHED
4416 * effective_load() calculates the load change as seen from the root_task_group
4418 * Adding load to a group doesn't make a group heavier, but can cause movement
4419 * of group shares between cpus. Assuming the shares were perfectly aligned one
4420 * can calculate the shift in shares.
4422 * Calculate the effective load difference if @wl is added (subtracted) to @tg
4423 * on this @cpu and results in a total addition (subtraction) of @wg to the
4424 * total group weight.
4426 * Given a runqueue weight distribution (rw_i) we can compute a shares
4427 * distribution (s_i) using:
4429 * s_i = rw_i / \Sum rw_j (1)
4431 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4432 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4433 * shares distribution (s_i):
4435 * rw_i = { 2, 4, 1, 0 }
4436 * s_i = { 2/7, 4/7, 1/7, 0 }
4438 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4439 * task used to run on and the CPU the waker is running on), we need to
4440 * compute the effect of waking a task on either CPU and, in case of a sync
4441 * wakeup, compute the effect of the current task going to sleep.
4443 * So for a change of @wl to the local @cpu with an overall group weight change
4444 * of @wl we can compute the new shares distribution (s'_i) using:
4446 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
4448 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4449 * differences in waking a task to CPU 0. The additional task changes the
4450 * weight and shares distributions like:
4452 * rw'_i = { 3, 4, 1, 0 }
4453 * s'_i = { 3/8, 4/8, 1/8, 0 }
4455 * We can then compute the difference in effective weight by using:
4457 * dw_i = S * (s'_i - s_i) (3)
4459 * Where 'S' is the group weight as seen by its parent.
4461 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4462 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4463 * 4/7) times the weight of the group.
4465 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4467 struct sched_entity *se = tg->se[cpu];
4469 if (!tg->parent) /* the trivial, non-cgroup case */
4472 for_each_sched_entity(se) {
4478 * W = @wg + \Sum rw_j
4480 W = wg + calc_tg_weight(tg, se->my_q);
4485 w = se->my_q->avg.load_avg + wl;
4488 * wl = S * s'_i; see (2)
4491 wl = (w * (long)tg->shares) / W;
4496 * Per the above, wl is the new se->load.weight value; since
4497 * those are clipped to [MIN_SHARES, ...) do so now. See
4498 * calc_cfs_shares().
4500 if (wl < MIN_SHARES)
4504 * wl = dw_i = S * (s'_i - s_i); see (3)
4506 wl -= se->avg.load_avg;
4509 * Recursively apply this logic to all parent groups to compute
4510 * the final effective load change on the root group. Since
4511 * only the @tg group gets extra weight, all parent groups can
4512 * only redistribute existing shares. @wl is the shift in shares
4513 * resulting from this level per the above.
4522 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4530 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
4531 * A waker of many should wake a different task than the one last awakened
4532 * at a frequency roughly N times higher than one of its wakees. In order
4533 * to determine whether we should let the load spread vs consolodating to
4534 * shared cache, we look for a minimum 'flip' frequency of llc_size in one
4535 * partner, and a factor of lls_size higher frequency in the other. With
4536 * both conditions met, we can be relatively sure that the relationship is
4537 * non-monogamous, with partner count exceeding socket size. Waker/wakee
4538 * being client/server, worker/dispatcher, interrupt source or whatever is
4539 * irrelevant, spread criteria is apparent partner count exceeds socket size.
4541 static int wake_wide(struct task_struct *p)
4543 unsigned int master = current->wakee_flips;
4544 unsigned int slave = p->wakee_flips;
4545 int factor = this_cpu_read(sd_llc_size);
4548 swap(master, slave);
4549 if (slave < factor || master < slave * factor)
4554 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4556 s64 this_load, load;
4557 s64 this_eff_load, prev_eff_load;
4558 int idx, this_cpu, prev_cpu;
4559 struct task_group *tg;
4560 unsigned long weight;
4564 this_cpu = smp_processor_id();
4565 prev_cpu = task_cpu(p);
4566 load = source_load(prev_cpu, idx);
4567 this_load = target_load(this_cpu, idx);
4570 * If sync wakeup then subtract the (maximum possible)
4571 * effect of the currently running task from the load
4572 * of the current CPU:
4575 tg = task_group(current);
4576 weight = current->se.avg.load_avg;
4578 this_load += effective_load(tg, this_cpu, -weight, -weight);
4579 load += effective_load(tg, prev_cpu, 0, -weight);
4583 weight = p->se.avg.load_avg;
4586 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4587 * due to the sync cause above having dropped this_load to 0, we'll
4588 * always have an imbalance, but there's really nothing you can do
4589 * about that, so that's good too.
4591 * Otherwise check if either cpus are near enough in load to allow this
4592 * task to be woken on this_cpu.
4594 this_eff_load = 100;
4595 this_eff_load *= capacity_of(prev_cpu);
4597 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4598 prev_eff_load *= capacity_of(this_cpu);
4600 if (this_load > 0) {
4601 this_eff_load *= this_load +
4602 effective_load(tg, this_cpu, weight, weight);
4604 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4607 balanced = this_eff_load <= prev_eff_load;
4609 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4614 schedstat_inc(sd, ttwu_move_affine);
4615 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4621 * find_idlest_group finds and returns the least busy CPU group within the
4624 static struct sched_group *
4625 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4626 int this_cpu, int sd_flag)
4628 struct sched_group *idlest = NULL, *group = sd->groups;
4629 unsigned long min_load = ULONG_MAX, this_load = 0;
4630 int load_idx = sd->forkexec_idx;
4631 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4633 if (sd_flag & SD_BALANCE_WAKE)
4634 load_idx = sd->wake_idx;
4637 unsigned long load, avg_load;
4641 /* Skip over this group if it has no CPUs allowed */
4642 if (!cpumask_intersects(sched_group_cpus(group),
4643 tsk_cpus_allowed(p)))
4646 local_group = cpumask_test_cpu(this_cpu,
4647 sched_group_cpus(group));
4649 /* Tally up the load of all CPUs in the group */
4652 for_each_cpu(i, sched_group_cpus(group)) {
4653 /* Bias balancing toward cpus of our domain */
4655 load = source_load(i, load_idx);
4657 load = target_load(i, load_idx);
4662 /* Adjust by relative CPU capacity of the group */
4663 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
4666 this_load = avg_load;
4667 } else if (avg_load < min_load) {
4668 min_load = avg_load;
4671 } while (group = group->next, group != sd->groups);
4673 if (!idlest || 100*this_load < imbalance*min_load)
4679 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4682 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4684 unsigned long load, min_load = ULONG_MAX;
4685 unsigned int min_exit_latency = UINT_MAX;
4686 u64 latest_idle_timestamp = 0;
4687 int least_loaded_cpu = this_cpu;
4688 int shallowest_idle_cpu = -1;
4691 /* Traverse only the allowed CPUs */
4692 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4694 struct rq *rq = cpu_rq(i);
4695 struct cpuidle_state *idle = idle_get_state(rq);
4696 if (idle && idle->exit_latency < min_exit_latency) {
4698 * We give priority to a CPU whose idle state
4699 * has the smallest exit latency irrespective
4700 * of any idle timestamp.
4702 min_exit_latency = idle->exit_latency;
4703 latest_idle_timestamp = rq->idle_stamp;
4704 shallowest_idle_cpu = i;
4705 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
4706 rq->idle_stamp > latest_idle_timestamp) {
4708 * If equal or no active idle state, then
4709 * the most recently idled CPU might have
4712 latest_idle_timestamp = rq->idle_stamp;
4713 shallowest_idle_cpu = i;
4715 } else if (shallowest_idle_cpu == -1) {
4716 load = weighted_cpuload(i);
4717 if (load < min_load || (load == min_load && i == this_cpu)) {
4719 least_loaded_cpu = i;
4724 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
4728 * Try and locate an idle CPU in the sched_domain.
4730 static int select_idle_sibling(struct task_struct *p, int target)
4732 struct sched_domain *sd;
4733 struct sched_group *sg;
4734 int i = task_cpu(p);
4736 if (idle_cpu(target))
4740 * If the prevous cpu is cache affine and idle, don't be stupid.
4742 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4746 * Otherwise, iterate the domains and find an elegible idle cpu.
4748 sd = rcu_dereference(per_cpu(sd_llc, target));
4749 for_each_lower_domain(sd) {
4752 if (!cpumask_intersects(sched_group_cpus(sg),
4753 tsk_cpus_allowed(p)))
4756 for_each_cpu(i, sched_group_cpus(sg)) {
4757 if (i == target || !idle_cpu(i))
4761 target = cpumask_first_and(sched_group_cpus(sg),
4762 tsk_cpus_allowed(p));
4766 } while (sg != sd->groups);
4772 * get_cpu_usage returns the amount of capacity of a CPU that is used by CFS
4773 * tasks. The unit of the return value must be the one of capacity so we can
4774 * compare the usage with the capacity of the CPU that is available for CFS
4775 * task (ie cpu_capacity).
4776 * cfs.avg.util_avg is the sum of running time of runnable tasks on a
4777 * CPU. It represents the amount of utilization of a CPU in the range
4778 * [0..SCHED_LOAD_SCALE]. The usage of a CPU can't be higher than the full
4779 * capacity of the CPU because it's about the running time on this CPU.
4780 * Nevertheless, cfs.avg.util_avg can be higher than SCHED_LOAD_SCALE
4781 * because of unfortunate rounding in util_avg or just
4782 * after migrating tasks until the average stabilizes with the new running
4783 * time. So we need to check that the usage stays into the range
4784 * [0..cpu_capacity_orig] and cap if necessary.
4785 * Without capping the usage, a group could be seen as overloaded (CPU0 usage
4786 * at 121% + CPU1 usage at 80%) whereas CPU1 has 20% of available capacity
4788 static int get_cpu_usage(int cpu)
4790 unsigned long usage = cpu_rq(cpu)->cfs.avg.util_avg;
4791 unsigned long capacity = capacity_orig_of(cpu);
4793 if (usage >= SCHED_LOAD_SCALE)
4796 return (usage * capacity) >> SCHED_LOAD_SHIFT;
4800 * select_task_rq_fair: Select target runqueue for the waking task in domains
4801 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
4802 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
4804 * Balances load by selecting the idlest cpu in the idlest group, or under
4805 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
4807 * Returns the target cpu number.
4809 * preempt must be disabled.
4812 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4814 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4815 int cpu = smp_processor_id();
4816 int new_cpu = prev_cpu;
4817 int want_affine = 0;
4818 int sync = wake_flags & WF_SYNC;
4820 if (sd_flag & SD_BALANCE_WAKE)
4821 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
4824 for_each_domain(cpu, tmp) {
4825 if (!(tmp->flags & SD_LOAD_BALANCE))
4829 * If both cpu and prev_cpu are part of this domain,
4830 * cpu is a valid SD_WAKE_AFFINE target.
4832 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4833 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4838 if (tmp->flags & sd_flag)
4840 else if (!want_affine)
4845 sd = NULL; /* Prefer wake_affine over balance flags */
4846 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4851 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
4852 new_cpu = select_idle_sibling(p, new_cpu);
4855 struct sched_group *group;
4858 if (!(sd->flags & sd_flag)) {
4863 group = find_idlest_group(sd, p, cpu, sd_flag);
4869 new_cpu = find_idlest_cpu(group, p, cpu);
4870 if (new_cpu == -1 || new_cpu == cpu) {
4871 /* Now try balancing at a lower domain level of cpu */
4876 /* Now try balancing at a lower domain level of new_cpu */
4878 weight = sd->span_weight;
4880 for_each_domain(cpu, tmp) {
4881 if (weight <= tmp->span_weight)
4883 if (tmp->flags & sd_flag)
4886 /* while loop will break here if sd == NULL */
4894 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4895 * cfs_rq_of(p) references at time of call are still valid and identify the
4896 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4897 * other assumptions, including the state of rq->lock, should be made.
4899 static void migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4902 * We are supposed to update the task to "current" time, then its up to date
4903 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
4904 * what current time is, so simply throw away the out-of-date time. This
4905 * will result in the wakee task is less decayed, but giving the wakee more
4906 * load sounds not bad.
4908 remove_entity_load_avg(&p->se);
4910 /* Tell new CPU we are migrated */
4911 p->se.avg.last_update_time = 0;
4913 /* We have migrated, no longer consider this task hot */
4914 p->se.exec_start = 0;
4916 #endif /* CONFIG_SMP */
4918 static unsigned long
4919 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4921 unsigned long gran = sysctl_sched_wakeup_granularity;
4924 * Since its curr running now, convert the gran from real-time
4925 * to virtual-time in his units.
4927 * By using 'se' instead of 'curr' we penalize light tasks, so
4928 * they get preempted easier. That is, if 'se' < 'curr' then
4929 * the resulting gran will be larger, therefore penalizing the
4930 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4931 * be smaller, again penalizing the lighter task.
4933 * This is especially important for buddies when the leftmost
4934 * task is higher priority than the buddy.
4936 return calc_delta_fair(gran, se);
4940 * Should 'se' preempt 'curr'.
4954 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4956 s64 gran, vdiff = curr->vruntime - se->vruntime;
4961 gran = wakeup_gran(curr, se);
4968 static void set_last_buddy(struct sched_entity *se)
4970 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4973 for_each_sched_entity(se)
4974 cfs_rq_of(se)->last = se;
4977 static void set_next_buddy(struct sched_entity *se)
4979 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4982 for_each_sched_entity(se)
4983 cfs_rq_of(se)->next = se;
4986 static void set_skip_buddy(struct sched_entity *se)
4988 for_each_sched_entity(se)
4989 cfs_rq_of(se)->skip = se;
4993 * Preempt the current task with a newly woken task if needed:
4995 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4997 struct task_struct *curr = rq->curr;
4998 struct sched_entity *se = &curr->se, *pse = &p->se;
4999 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5000 int scale = cfs_rq->nr_running >= sched_nr_latency;
5001 int next_buddy_marked = 0;
5003 if (unlikely(se == pse))
5007 * This is possible from callers such as attach_tasks(), in which we
5008 * unconditionally check_prempt_curr() after an enqueue (which may have
5009 * lead to a throttle). This both saves work and prevents false
5010 * next-buddy nomination below.
5012 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5015 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5016 set_next_buddy(pse);
5017 next_buddy_marked = 1;
5021 * We can come here with TIF_NEED_RESCHED already set from new task
5024 * Note: this also catches the edge-case of curr being in a throttled
5025 * group (e.g. via set_curr_task), since update_curr() (in the
5026 * enqueue of curr) will have resulted in resched being set. This
5027 * prevents us from potentially nominating it as a false LAST_BUDDY
5030 if (test_tsk_need_resched(curr))
5033 /* Idle tasks are by definition preempted by non-idle tasks. */
5034 if (unlikely(curr->policy == SCHED_IDLE) &&
5035 likely(p->policy != SCHED_IDLE))
5039 * Batch and idle tasks do not preempt non-idle tasks (their preemption
5040 * is driven by the tick):
5042 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5045 find_matching_se(&se, &pse);
5046 update_curr(cfs_rq_of(se));
5048 if (wakeup_preempt_entity(se, pse) == 1) {
5050 * Bias pick_next to pick the sched entity that is
5051 * triggering this preemption.
5053 if (!next_buddy_marked)
5054 set_next_buddy(pse);
5063 * Only set the backward buddy when the current task is still
5064 * on the rq. This can happen when a wakeup gets interleaved
5065 * with schedule on the ->pre_schedule() or idle_balance()
5066 * point, either of which can * drop the rq lock.
5068 * Also, during early boot the idle thread is in the fair class,
5069 * for obvious reasons its a bad idea to schedule back to it.
5071 if (unlikely(!se->on_rq || curr == rq->idle))
5074 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5078 static struct task_struct *
5079 pick_next_task_fair(struct rq *rq, struct task_struct *prev)
5081 struct cfs_rq *cfs_rq = &rq->cfs;
5082 struct sched_entity *se;
5083 struct task_struct *p;
5087 #ifdef CONFIG_FAIR_GROUP_SCHED
5088 if (!cfs_rq->nr_running)
5091 if (prev->sched_class != &fair_sched_class)
5095 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5096 * likely that a next task is from the same cgroup as the current.
5098 * Therefore attempt to avoid putting and setting the entire cgroup
5099 * hierarchy, only change the part that actually changes.
5103 struct sched_entity *curr = cfs_rq->curr;
5106 * Since we got here without doing put_prev_entity() we also
5107 * have to consider cfs_rq->curr. If it is still a runnable
5108 * entity, update_curr() will update its vruntime, otherwise
5109 * forget we've ever seen it.
5113 update_curr(cfs_rq);
5118 * This call to check_cfs_rq_runtime() will do the
5119 * throttle and dequeue its entity in the parent(s).
5120 * Therefore the 'simple' nr_running test will indeed
5123 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5127 se = pick_next_entity(cfs_rq, curr);
5128 cfs_rq = group_cfs_rq(se);
5134 * Since we haven't yet done put_prev_entity and if the selected task
5135 * is a different task than we started out with, try and touch the
5136 * least amount of cfs_rqs.
5139 struct sched_entity *pse = &prev->se;
5141 while (!(cfs_rq = is_same_group(se, pse))) {
5142 int se_depth = se->depth;
5143 int pse_depth = pse->depth;
5145 if (se_depth <= pse_depth) {
5146 put_prev_entity(cfs_rq_of(pse), pse);
5147 pse = parent_entity(pse);
5149 if (se_depth >= pse_depth) {
5150 set_next_entity(cfs_rq_of(se), se);
5151 se = parent_entity(se);
5155 put_prev_entity(cfs_rq, pse);
5156 set_next_entity(cfs_rq, se);
5159 if (hrtick_enabled(rq))
5160 hrtick_start_fair(rq, p);
5167 if (!cfs_rq->nr_running)
5170 put_prev_task(rq, prev);
5173 se = pick_next_entity(cfs_rq, NULL);
5174 set_next_entity(cfs_rq, se);
5175 cfs_rq = group_cfs_rq(se);
5180 if (hrtick_enabled(rq))
5181 hrtick_start_fair(rq, p);
5187 * This is OK, because current is on_cpu, which avoids it being picked
5188 * for load-balance and preemption/IRQs are still disabled avoiding
5189 * further scheduler activity on it and we're being very careful to
5190 * re-start the picking loop.
5192 lockdep_unpin_lock(&rq->lock);
5193 new_tasks = idle_balance(rq);
5194 lockdep_pin_lock(&rq->lock);
5196 * Because idle_balance() releases (and re-acquires) rq->lock, it is
5197 * possible for any higher priority task to appear. In that case we
5198 * must re-start the pick_next_entity() loop.
5210 * Account for a descheduled task:
5212 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5214 struct sched_entity *se = &prev->se;
5215 struct cfs_rq *cfs_rq;
5217 for_each_sched_entity(se) {
5218 cfs_rq = cfs_rq_of(se);
5219 put_prev_entity(cfs_rq, se);
5224 * sched_yield() is very simple
5226 * The magic of dealing with the ->skip buddy is in pick_next_entity.
5228 static void yield_task_fair(struct rq *rq)
5230 struct task_struct *curr = rq->curr;
5231 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5232 struct sched_entity *se = &curr->se;
5235 * Are we the only task in the tree?
5237 if (unlikely(rq->nr_running == 1))
5240 clear_buddies(cfs_rq, se);
5242 if (curr->policy != SCHED_BATCH) {
5243 update_rq_clock(rq);
5245 * Update run-time statistics of the 'current'.
5247 update_curr(cfs_rq);
5249 * Tell update_rq_clock() that we've just updated,
5250 * so we don't do microscopic update in schedule()
5251 * and double the fastpath cost.
5253 rq_clock_skip_update(rq, true);
5259 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5261 struct sched_entity *se = &p->se;
5263 /* throttled hierarchies are not runnable */
5264 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5267 /* Tell the scheduler that we'd really like pse to run next. */
5270 yield_task_fair(rq);
5276 /**************************************************
5277 * Fair scheduling class load-balancing methods.
5281 * The purpose of load-balancing is to achieve the same basic fairness the
5282 * per-cpu scheduler provides, namely provide a proportional amount of compute
5283 * time to each task. This is expressed in the following equation:
5285 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
5287 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5288 * W_i,0 is defined as:
5290 * W_i,0 = \Sum_j w_i,j (2)
5292 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5293 * is derived from the nice value as per prio_to_weight[].
5295 * The weight average is an exponential decay average of the instantaneous
5298 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
5300 * C_i is the compute capacity of cpu i, typically it is the
5301 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5302 * can also include other factors [XXX].
5304 * To achieve this balance we define a measure of imbalance which follows
5305 * directly from (1):
5307 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
5309 * We them move tasks around to minimize the imbalance. In the continuous
5310 * function space it is obvious this converges, in the discrete case we get
5311 * a few fun cases generally called infeasible weight scenarios.
5314 * - infeasible weights;
5315 * - local vs global optima in the discrete case. ]
5320 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5321 * for all i,j solution, we create a tree of cpus that follows the hardware
5322 * topology where each level pairs two lower groups (or better). This results
5323 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5324 * tree to only the first of the previous level and we decrease the frequency
5325 * of load-balance at each level inv. proportional to the number of cpus in
5331 * \Sum { --- * --- * 2^i } = O(n) (5)
5333 * `- size of each group
5334 * | | `- number of cpus doing load-balance
5336 * `- sum over all levels
5338 * Coupled with a limit on how many tasks we can migrate every balance pass,
5339 * this makes (5) the runtime complexity of the balancer.
5341 * An important property here is that each CPU is still (indirectly) connected
5342 * to every other cpu in at most O(log n) steps:
5344 * The adjacency matrix of the resulting graph is given by:
5347 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
5350 * And you'll find that:
5352 * A^(log_2 n)_i,j != 0 for all i,j (7)
5354 * Showing there's indeed a path between every cpu in at most O(log n) steps.
5355 * The task movement gives a factor of O(m), giving a convergence complexity
5358 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
5363 * In order to avoid CPUs going idle while there's still work to do, new idle
5364 * balancing is more aggressive and has the newly idle cpu iterate up the domain
5365 * tree itself instead of relying on other CPUs to bring it work.
5367 * This adds some complexity to both (5) and (8) but it reduces the total idle
5375 * Cgroups make a horror show out of (2), instead of a simple sum we get:
5378 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
5383 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
5385 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5387 * The big problem is S_k, its a global sum needed to compute a local (W_i)
5390 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5391 * rewrite all of this once again.]
5394 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5396 enum fbq_type { regular, remote, all };
5398 #define LBF_ALL_PINNED 0x01
5399 #define LBF_NEED_BREAK 0x02
5400 #define LBF_DST_PINNED 0x04
5401 #define LBF_SOME_PINNED 0x08
5404 struct sched_domain *sd;
5412 struct cpumask *dst_grpmask;
5414 enum cpu_idle_type idle;
5416 /* The set of CPUs under consideration for load-balancing */
5417 struct cpumask *cpus;
5422 unsigned int loop_break;
5423 unsigned int loop_max;
5425 enum fbq_type fbq_type;
5426 struct list_head tasks;
5430 * Is this task likely cache-hot:
5432 static int task_hot(struct task_struct *p, struct lb_env *env)
5436 lockdep_assert_held(&env->src_rq->lock);
5438 if (p->sched_class != &fair_sched_class)
5441 if (unlikely(p->policy == SCHED_IDLE))
5445 * Buddy candidates are cache hot:
5447 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
5448 (&p->se == cfs_rq_of(&p->se)->next ||
5449 &p->se == cfs_rq_of(&p->se)->last))
5452 if (sysctl_sched_migration_cost == -1)
5454 if (sysctl_sched_migration_cost == 0)
5457 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
5459 return delta < (s64)sysctl_sched_migration_cost;
5462 #ifdef CONFIG_NUMA_BALANCING
5464 * Returns 1, if task migration degrades locality
5465 * Returns 0, if task migration improves locality i.e migration preferred.
5466 * Returns -1, if task migration is not affected by locality.
5468 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
5470 struct numa_group *numa_group = rcu_dereference(p->numa_group);
5471 unsigned long src_faults, dst_faults;
5472 int src_nid, dst_nid;
5474 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
5477 if (!sched_feat(NUMA))
5480 src_nid = cpu_to_node(env->src_cpu);
5481 dst_nid = cpu_to_node(env->dst_cpu);
5483 if (src_nid == dst_nid)
5486 /* Migrating away from the preferred node is always bad. */
5487 if (src_nid == p->numa_preferred_nid) {
5488 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
5494 /* Encourage migration to the preferred node. */
5495 if (dst_nid == p->numa_preferred_nid)
5499 src_faults = group_faults(p, src_nid);
5500 dst_faults = group_faults(p, dst_nid);
5502 src_faults = task_faults(p, src_nid);
5503 dst_faults = task_faults(p, dst_nid);
5506 return dst_faults < src_faults;
5510 static inline int migrate_degrades_locality(struct task_struct *p,
5518 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
5521 int can_migrate_task(struct task_struct *p, struct lb_env *env)
5525 lockdep_assert_held(&env->src_rq->lock);
5528 * We do not migrate tasks that are:
5529 * 1) throttled_lb_pair, or
5530 * 2) cannot be migrated to this CPU due to cpus_allowed, or
5531 * 3) running (obviously), or
5532 * 4) are cache-hot on their current CPU.
5534 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
5537 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
5540 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
5542 env->flags |= LBF_SOME_PINNED;
5545 * Remember if this task can be migrated to any other cpu in
5546 * our sched_group. We may want to revisit it if we couldn't
5547 * meet load balance goals by pulling other tasks on src_cpu.
5549 * Also avoid computing new_dst_cpu if we have already computed
5550 * one in current iteration.
5552 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
5555 /* Prevent to re-select dst_cpu via env's cpus */
5556 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
5557 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
5558 env->flags |= LBF_DST_PINNED;
5559 env->new_dst_cpu = cpu;
5567 /* Record that we found atleast one task that could run on dst_cpu */
5568 env->flags &= ~LBF_ALL_PINNED;
5570 if (task_running(env->src_rq, p)) {
5571 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
5576 * Aggressive migration if:
5577 * 1) destination numa is preferred
5578 * 2) task is cache cold, or
5579 * 3) too many balance attempts have failed.
5581 tsk_cache_hot = migrate_degrades_locality(p, env);
5582 if (tsk_cache_hot == -1)
5583 tsk_cache_hot = task_hot(p, env);
5585 if (tsk_cache_hot <= 0 ||
5586 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
5587 if (tsk_cache_hot == 1) {
5588 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
5589 schedstat_inc(p, se.statistics.nr_forced_migrations);
5594 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
5599 * detach_task() -- detach the task for the migration specified in env
5601 static void detach_task(struct task_struct *p, struct lb_env *env)
5603 lockdep_assert_held(&env->src_rq->lock);
5605 deactivate_task(env->src_rq, p, 0);
5606 p->on_rq = TASK_ON_RQ_MIGRATING;
5607 set_task_cpu(p, env->dst_cpu);
5611 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
5612 * part of active balancing operations within "domain".
5614 * Returns a task if successful and NULL otherwise.
5616 static struct task_struct *detach_one_task(struct lb_env *env)
5618 struct task_struct *p, *n;
5620 lockdep_assert_held(&env->src_rq->lock);
5622 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
5623 if (!can_migrate_task(p, env))
5626 detach_task(p, env);
5629 * Right now, this is only the second place where
5630 * lb_gained[env->idle] is updated (other is detach_tasks)
5631 * so we can safely collect stats here rather than
5632 * inside detach_tasks().
5634 schedstat_inc(env->sd, lb_gained[env->idle]);
5640 static const unsigned int sched_nr_migrate_break = 32;
5643 * detach_tasks() -- tries to detach up to imbalance weighted load from
5644 * busiest_rq, as part of a balancing operation within domain "sd".
5646 * Returns number of detached tasks if successful and 0 otherwise.
5648 static int detach_tasks(struct lb_env *env)
5650 struct list_head *tasks = &env->src_rq->cfs_tasks;
5651 struct task_struct *p;
5655 lockdep_assert_held(&env->src_rq->lock);
5657 if (env->imbalance <= 0)
5660 while (!list_empty(tasks)) {
5662 * We don't want to steal all, otherwise we may be treated likewise,
5663 * which could at worst lead to a livelock crash.
5665 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
5668 p = list_first_entry(tasks, struct task_struct, se.group_node);
5671 /* We've more or less seen every task there is, call it quits */
5672 if (env->loop > env->loop_max)
5675 /* take a breather every nr_migrate tasks */
5676 if (env->loop > env->loop_break) {
5677 env->loop_break += sched_nr_migrate_break;
5678 env->flags |= LBF_NEED_BREAK;
5682 if (!can_migrate_task(p, env))
5685 load = task_h_load(p);
5687 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5690 if ((load / 2) > env->imbalance)
5693 detach_task(p, env);
5694 list_add(&p->se.group_node, &env->tasks);
5697 env->imbalance -= load;
5699 #ifdef CONFIG_PREEMPT
5701 * NEWIDLE balancing is a source of latency, so preemptible
5702 * kernels will stop after the first task is detached to minimize
5703 * the critical section.
5705 if (env->idle == CPU_NEWLY_IDLE)
5710 * We only want to steal up to the prescribed amount of
5713 if (env->imbalance <= 0)
5718 list_move_tail(&p->se.group_node, tasks);
5722 * Right now, this is one of only two places we collect this stat
5723 * so we can safely collect detach_one_task() stats here rather
5724 * than inside detach_one_task().
5726 schedstat_add(env->sd, lb_gained[env->idle], detached);
5732 * attach_task() -- attach the task detached by detach_task() to its new rq.
5734 static void attach_task(struct rq *rq, struct task_struct *p)
5736 lockdep_assert_held(&rq->lock);
5738 BUG_ON(task_rq(p) != rq);
5739 p->on_rq = TASK_ON_RQ_QUEUED;
5740 activate_task(rq, p, 0);
5741 check_preempt_curr(rq, p, 0);
5745 * attach_one_task() -- attaches the task returned from detach_one_task() to
5748 static void attach_one_task(struct rq *rq, struct task_struct *p)
5750 raw_spin_lock(&rq->lock);
5752 raw_spin_unlock(&rq->lock);
5756 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
5759 static void attach_tasks(struct lb_env *env)
5761 struct list_head *tasks = &env->tasks;
5762 struct task_struct *p;
5764 raw_spin_lock(&env->dst_rq->lock);
5766 while (!list_empty(tasks)) {
5767 p = list_first_entry(tasks, struct task_struct, se.group_node);
5768 list_del_init(&p->se.group_node);
5770 attach_task(env->dst_rq, p);
5773 raw_spin_unlock(&env->dst_rq->lock);
5776 #ifdef CONFIG_FAIR_GROUP_SCHED
5777 static void update_blocked_averages(int cpu)
5779 struct rq *rq = cpu_rq(cpu);
5780 struct cfs_rq *cfs_rq;
5781 unsigned long flags;
5783 raw_spin_lock_irqsave(&rq->lock, flags);
5784 update_rq_clock(rq);
5787 * Iterates the task_group tree in a bottom up fashion, see
5788 * list_add_leaf_cfs_rq() for details.
5790 for_each_leaf_cfs_rq(rq, cfs_rq) {
5791 /* throttled entities do not contribute to load */
5792 if (throttled_hierarchy(cfs_rq))
5795 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
5796 update_tg_load_avg(cfs_rq, 0);
5798 raw_spin_unlock_irqrestore(&rq->lock, flags);
5802 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5803 * This needs to be done in a top-down fashion because the load of a child
5804 * group is a fraction of its parents load.
5806 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5808 struct rq *rq = rq_of(cfs_rq);
5809 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5810 unsigned long now = jiffies;
5813 if (cfs_rq->last_h_load_update == now)
5816 cfs_rq->h_load_next = NULL;
5817 for_each_sched_entity(se) {
5818 cfs_rq = cfs_rq_of(se);
5819 cfs_rq->h_load_next = se;
5820 if (cfs_rq->last_h_load_update == now)
5825 cfs_rq->h_load = cfs_rq->avg.load_avg;
5826 cfs_rq->last_h_load_update = now;
5829 while ((se = cfs_rq->h_load_next) != NULL) {
5830 load = cfs_rq->h_load;
5831 load = div64_ul(load * se->avg.load_avg, cfs_rq->avg.load_avg + 1);
5832 cfs_rq = group_cfs_rq(se);
5833 cfs_rq->h_load = load;
5834 cfs_rq->last_h_load_update = now;
5838 static unsigned long task_h_load(struct task_struct *p)
5840 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5842 update_cfs_rq_h_load(cfs_rq);
5843 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
5844 cfs_rq->avg.load_avg + 1);
5847 static inline void update_blocked_averages(int cpu)
5851 static unsigned long task_h_load(struct task_struct *p)
5853 return p->se.avg.load_avg;
5857 /********** Helpers for find_busiest_group ************************/
5866 * sg_lb_stats - stats of a sched_group required for load_balancing
5868 struct sg_lb_stats {
5869 unsigned long avg_load; /*Avg load across the CPUs of the group */
5870 unsigned long group_load; /* Total load over the CPUs of the group */
5871 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5872 unsigned long load_per_task;
5873 unsigned long group_capacity;
5874 unsigned long group_usage; /* Total usage of the group */
5875 unsigned int sum_nr_running; /* Nr tasks running in the group */
5876 unsigned int idle_cpus;
5877 unsigned int group_weight;
5878 enum group_type group_type;
5879 int group_no_capacity;
5880 #ifdef CONFIG_NUMA_BALANCING
5881 unsigned int nr_numa_running;
5882 unsigned int nr_preferred_running;
5887 * sd_lb_stats - Structure to store the statistics of a sched_domain
5888 * during load balancing.
5890 struct sd_lb_stats {
5891 struct sched_group *busiest; /* Busiest group in this sd */
5892 struct sched_group *local; /* Local group in this sd */
5893 unsigned long total_load; /* Total load of all groups in sd */
5894 unsigned long total_capacity; /* Total capacity of all groups in sd */
5895 unsigned long avg_load; /* Average load across all groups in sd */
5897 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5898 struct sg_lb_stats local_stat; /* Statistics of the local group */
5901 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5904 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5905 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5906 * We must however clear busiest_stat::avg_load because
5907 * update_sd_pick_busiest() reads this before assignment.
5909 *sds = (struct sd_lb_stats){
5913 .total_capacity = 0UL,
5916 .sum_nr_running = 0,
5917 .group_type = group_other,
5923 * get_sd_load_idx - Obtain the load index for a given sched domain.
5924 * @sd: The sched_domain whose load_idx is to be obtained.
5925 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5927 * Return: The load index.
5929 static inline int get_sd_load_idx(struct sched_domain *sd,
5930 enum cpu_idle_type idle)
5936 load_idx = sd->busy_idx;
5939 case CPU_NEWLY_IDLE:
5940 load_idx = sd->newidle_idx;
5943 load_idx = sd->idle_idx;
5950 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5952 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1))
5953 return sd->smt_gain / sd->span_weight;
5955 return SCHED_CAPACITY_SCALE;
5958 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
5960 return default_scale_cpu_capacity(sd, cpu);
5963 static unsigned long scale_rt_capacity(int cpu)
5965 struct rq *rq = cpu_rq(cpu);
5966 u64 total, used, age_stamp, avg;
5970 * Since we're reading these variables without serialization make sure
5971 * we read them once before doing sanity checks on them.
5973 age_stamp = READ_ONCE(rq->age_stamp);
5974 avg = READ_ONCE(rq->rt_avg);
5975 delta = __rq_clock_broken(rq) - age_stamp;
5977 if (unlikely(delta < 0))
5980 total = sched_avg_period() + delta;
5982 used = div_u64(avg, total);
5984 if (likely(used < SCHED_CAPACITY_SCALE))
5985 return SCHED_CAPACITY_SCALE - used;
5990 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
5992 unsigned long capacity = SCHED_CAPACITY_SCALE;
5993 struct sched_group *sdg = sd->groups;
5995 if (sched_feat(ARCH_CAPACITY))
5996 capacity *= arch_scale_cpu_capacity(sd, cpu);
5998 capacity *= default_scale_cpu_capacity(sd, cpu);
6000 capacity >>= SCHED_CAPACITY_SHIFT;
6002 cpu_rq(cpu)->cpu_capacity_orig = capacity;
6004 capacity *= scale_rt_capacity(cpu);
6005 capacity >>= SCHED_CAPACITY_SHIFT;
6010 cpu_rq(cpu)->cpu_capacity = capacity;
6011 sdg->sgc->capacity = capacity;
6014 void update_group_capacity(struct sched_domain *sd, int cpu)
6016 struct sched_domain *child = sd->child;
6017 struct sched_group *group, *sdg = sd->groups;
6018 unsigned long capacity;
6019 unsigned long interval;
6021 interval = msecs_to_jiffies(sd->balance_interval);
6022 interval = clamp(interval, 1UL, max_load_balance_interval);
6023 sdg->sgc->next_update = jiffies + interval;
6026 update_cpu_capacity(sd, cpu);
6032 if (child->flags & SD_OVERLAP) {
6034 * SD_OVERLAP domains cannot assume that child groups
6035 * span the current group.
6038 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6039 struct sched_group_capacity *sgc;
6040 struct rq *rq = cpu_rq(cpu);
6043 * build_sched_domains() -> init_sched_groups_capacity()
6044 * gets here before we've attached the domains to the
6047 * Use capacity_of(), which is set irrespective of domains
6048 * in update_cpu_capacity().
6050 * This avoids capacity from being 0 and
6051 * causing divide-by-zero issues on boot.
6053 if (unlikely(!rq->sd)) {
6054 capacity += capacity_of(cpu);
6058 sgc = rq->sd->groups->sgc;
6059 capacity += sgc->capacity;
6063 * !SD_OVERLAP domains can assume that child groups
6064 * span the current group.
6067 group = child->groups;
6069 capacity += group->sgc->capacity;
6070 group = group->next;
6071 } while (group != child->groups);
6074 sdg->sgc->capacity = capacity;
6078 * Check whether the capacity of the rq has been noticeably reduced by side
6079 * activity. The imbalance_pct is used for the threshold.
6080 * Return true is the capacity is reduced
6083 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6085 return ((rq->cpu_capacity * sd->imbalance_pct) <
6086 (rq->cpu_capacity_orig * 100));
6090 * Group imbalance indicates (and tries to solve) the problem where balancing
6091 * groups is inadequate due to tsk_cpus_allowed() constraints.
6093 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6094 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6097 * { 0 1 2 3 } { 4 5 6 7 }
6100 * If we were to balance group-wise we'd place two tasks in the first group and
6101 * two tasks in the second group. Clearly this is undesired as it will overload
6102 * cpu 3 and leave one of the cpus in the second group unused.
6104 * The current solution to this issue is detecting the skew in the first group
6105 * by noticing the lower domain failed to reach balance and had difficulty
6106 * moving tasks due to affinity constraints.
6108 * When this is so detected; this group becomes a candidate for busiest; see
6109 * update_sd_pick_busiest(). And calculate_imbalance() and
6110 * find_busiest_group() avoid some of the usual balance conditions to allow it
6111 * to create an effective group imbalance.
6113 * This is a somewhat tricky proposition since the next run might not find the
6114 * group imbalance and decide the groups need to be balanced again. A most
6115 * subtle and fragile situation.
6118 static inline int sg_imbalanced(struct sched_group *group)
6120 return group->sgc->imbalance;
6124 * group_has_capacity returns true if the group has spare capacity that could
6125 * be used by some tasks.
6126 * We consider that a group has spare capacity if the * number of task is
6127 * smaller than the number of CPUs or if the usage is lower than the available
6128 * capacity for CFS tasks.
6129 * For the latter, we use a threshold to stabilize the state, to take into
6130 * account the variance of the tasks' load and to return true if the available
6131 * capacity in meaningful for the load balancer.
6132 * As an example, an available capacity of 1% can appear but it doesn't make
6133 * any benefit for the load balance.
6136 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6138 if (sgs->sum_nr_running < sgs->group_weight)
6141 if ((sgs->group_capacity * 100) >
6142 (sgs->group_usage * env->sd->imbalance_pct))
6149 * group_is_overloaded returns true if the group has more tasks than it can
6151 * group_is_overloaded is not equals to !group_has_capacity because a group
6152 * with the exact right number of tasks, has no more spare capacity but is not
6153 * overloaded so both group_has_capacity and group_is_overloaded return
6157 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6159 if (sgs->sum_nr_running <= sgs->group_weight)
6162 if ((sgs->group_capacity * 100) <
6163 (sgs->group_usage * env->sd->imbalance_pct))
6169 static enum group_type group_classify(struct lb_env *env,
6170 struct sched_group *group,
6171 struct sg_lb_stats *sgs)
6173 if (sgs->group_no_capacity)
6174 return group_overloaded;
6176 if (sg_imbalanced(group))
6177 return group_imbalanced;
6183 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6184 * @env: The load balancing environment.
6185 * @group: sched_group whose statistics are to be updated.
6186 * @load_idx: Load index of sched_domain of this_cpu for load calc.
6187 * @local_group: Does group contain this_cpu.
6188 * @sgs: variable to hold the statistics for this group.
6189 * @overload: Indicate more than one runnable task for any CPU.
6191 static inline void update_sg_lb_stats(struct lb_env *env,
6192 struct sched_group *group, int load_idx,
6193 int local_group, struct sg_lb_stats *sgs,
6199 memset(sgs, 0, sizeof(*sgs));
6201 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6202 struct rq *rq = cpu_rq(i);
6204 /* Bias balancing toward cpus of our domain */
6206 load = target_load(i, load_idx);
6208 load = source_load(i, load_idx);
6210 sgs->group_load += load;
6211 sgs->group_usage += get_cpu_usage(i);
6212 sgs->sum_nr_running += rq->cfs.h_nr_running;
6214 if (rq->nr_running > 1)
6217 #ifdef CONFIG_NUMA_BALANCING
6218 sgs->nr_numa_running += rq->nr_numa_running;
6219 sgs->nr_preferred_running += rq->nr_preferred_running;
6221 sgs->sum_weighted_load += weighted_cpuload(i);
6226 /* Adjust by relative CPU capacity of the group */
6227 sgs->group_capacity = group->sgc->capacity;
6228 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6230 if (sgs->sum_nr_running)
6231 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6233 sgs->group_weight = group->group_weight;
6235 sgs->group_no_capacity = group_is_overloaded(env, sgs);
6236 sgs->group_type = group_classify(env, group, sgs);
6240 * update_sd_pick_busiest - return 1 on busiest group
6241 * @env: The load balancing environment.
6242 * @sds: sched_domain statistics
6243 * @sg: sched_group candidate to be checked for being the busiest
6244 * @sgs: sched_group statistics
6246 * Determine if @sg is a busier group than the previously selected
6249 * Return: %true if @sg is a busier group than the previously selected
6250 * busiest group. %false otherwise.
6252 static bool update_sd_pick_busiest(struct lb_env *env,
6253 struct sd_lb_stats *sds,
6254 struct sched_group *sg,
6255 struct sg_lb_stats *sgs)
6257 struct sg_lb_stats *busiest = &sds->busiest_stat;
6259 if (sgs->group_type > busiest->group_type)
6262 if (sgs->group_type < busiest->group_type)
6265 if (sgs->avg_load <= busiest->avg_load)
6268 /* This is the busiest node in its class. */
6269 if (!(env->sd->flags & SD_ASYM_PACKING))
6273 * ASYM_PACKING needs to move all the work to the lowest
6274 * numbered CPUs in the group, therefore mark all groups
6275 * higher than ourself as busy.
6277 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6281 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
6288 #ifdef CONFIG_NUMA_BALANCING
6289 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6291 if (sgs->sum_nr_running > sgs->nr_numa_running)
6293 if (sgs->sum_nr_running > sgs->nr_preferred_running)
6298 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6300 if (rq->nr_running > rq->nr_numa_running)
6302 if (rq->nr_running > rq->nr_preferred_running)
6307 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6312 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6316 #endif /* CONFIG_NUMA_BALANCING */
6319 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6320 * @env: The load balancing environment.
6321 * @sds: variable to hold the statistics for this sched_domain.
6323 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6325 struct sched_domain *child = env->sd->child;
6326 struct sched_group *sg = env->sd->groups;
6327 struct sg_lb_stats tmp_sgs;
6328 int load_idx, prefer_sibling = 0;
6329 bool overload = false;
6331 if (child && child->flags & SD_PREFER_SIBLING)
6334 load_idx = get_sd_load_idx(env->sd, env->idle);
6337 struct sg_lb_stats *sgs = &tmp_sgs;
6340 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6343 sgs = &sds->local_stat;
6345 if (env->idle != CPU_NEWLY_IDLE ||
6346 time_after_eq(jiffies, sg->sgc->next_update))
6347 update_group_capacity(env->sd, env->dst_cpu);
6350 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6357 * In case the child domain prefers tasks go to siblings
6358 * first, lower the sg capacity so that we'll try
6359 * and move all the excess tasks away. We lower the capacity
6360 * of a group only if the local group has the capacity to fit
6361 * these excess tasks. The extra check prevents the case where
6362 * you always pull from the heaviest group when it is already
6363 * under-utilized (possible with a large weight task outweighs
6364 * the tasks on the system).
6366 if (prefer_sibling && sds->local &&
6367 group_has_capacity(env, &sds->local_stat) &&
6368 (sgs->sum_nr_running > 1)) {
6369 sgs->group_no_capacity = 1;
6370 sgs->group_type = group_overloaded;
6373 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6375 sds->busiest_stat = *sgs;
6379 /* Now, start updating sd_lb_stats */
6380 sds->total_load += sgs->group_load;
6381 sds->total_capacity += sgs->group_capacity;
6384 } while (sg != env->sd->groups);
6386 if (env->sd->flags & SD_NUMA)
6387 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6389 if (!env->sd->parent) {
6390 /* update overload indicator if we are at root domain */
6391 if (env->dst_rq->rd->overload != overload)
6392 env->dst_rq->rd->overload = overload;
6398 * check_asym_packing - Check to see if the group is packed into the
6401 * This is primarily intended to used at the sibling level. Some
6402 * cores like POWER7 prefer to use lower numbered SMT threads. In the
6403 * case of POWER7, it can move to lower SMT modes only when higher
6404 * threads are idle. When in lower SMT modes, the threads will
6405 * perform better since they share less core resources. Hence when we
6406 * have idle threads, we want them to be the higher ones.
6408 * This packing function is run on idle threads. It checks to see if
6409 * the busiest CPU in this domain (core in the P7 case) has a higher
6410 * CPU number than the packing function is being run on. Here we are
6411 * assuming lower CPU number will be equivalent to lower a SMT thread
6414 * Return: 1 when packing is required and a task should be moved to
6415 * this CPU. The amount of the imbalance is returned in *imbalance.
6417 * @env: The load balancing environment.
6418 * @sds: Statistics of the sched_domain which is to be packed
6420 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
6424 if (!(env->sd->flags & SD_ASYM_PACKING))
6430 busiest_cpu = group_first_cpu(sds->busiest);
6431 if (env->dst_cpu > busiest_cpu)
6434 env->imbalance = DIV_ROUND_CLOSEST(
6435 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
6436 SCHED_CAPACITY_SCALE);
6442 * fix_small_imbalance - Calculate the minor imbalance that exists
6443 * amongst the groups of a sched_domain, during
6445 * @env: The load balancing environment.
6446 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
6449 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6451 unsigned long tmp, capa_now = 0, capa_move = 0;
6452 unsigned int imbn = 2;
6453 unsigned long scaled_busy_load_per_task;
6454 struct sg_lb_stats *local, *busiest;
6456 local = &sds->local_stat;
6457 busiest = &sds->busiest_stat;
6459 if (!local->sum_nr_running)
6460 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
6461 else if (busiest->load_per_task > local->load_per_task)
6464 scaled_busy_load_per_task =
6465 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6466 busiest->group_capacity;
6468 if (busiest->avg_load + scaled_busy_load_per_task >=
6469 local->avg_load + (scaled_busy_load_per_task * imbn)) {
6470 env->imbalance = busiest->load_per_task;
6475 * OK, we don't have enough imbalance to justify moving tasks,
6476 * however we may be able to increase total CPU capacity used by
6480 capa_now += busiest->group_capacity *
6481 min(busiest->load_per_task, busiest->avg_load);
6482 capa_now += local->group_capacity *
6483 min(local->load_per_task, local->avg_load);
6484 capa_now /= SCHED_CAPACITY_SCALE;
6486 /* Amount of load we'd subtract */
6487 if (busiest->avg_load > scaled_busy_load_per_task) {
6488 capa_move += busiest->group_capacity *
6489 min(busiest->load_per_task,
6490 busiest->avg_load - scaled_busy_load_per_task);
6493 /* Amount of load we'd add */
6494 if (busiest->avg_load * busiest->group_capacity <
6495 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
6496 tmp = (busiest->avg_load * busiest->group_capacity) /
6497 local->group_capacity;
6499 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
6500 local->group_capacity;
6502 capa_move += local->group_capacity *
6503 min(local->load_per_task, local->avg_load + tmp);
6504 capa_move /= SCHED_CAPACITY_SCALE;
6506 /* Move if we gain throughput */
6507 if (capa_move > capa_now)
6508 env->imbalance = busiest->load_per_task;
6512 * calculate_imbalance - Calculate the amount of imbalance present within the
6513 * groups of a given sched_domain during load balance.
6514 * @env: load balance environment
6515 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
6517 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
6519 unsigned long max_pull, load_above_capacity = ~0UL;
6520 struct sg_lb_stats *local, *busiest;
6522 local = &sds->local_stat;
6523 busiest = &sds->busiest_stat;
6525 if (busiest->group_type == group_imbalanced) {
6527 * In the group_imb case we cannot rely on group-wide averages
6528 * to ensure cpu-load equilibrium, look at wider averages. XXX
6530 busiest->load_per_task =
6531 min(busiest->load_per_task, sds->avg_load);
6535 * In the presence of smp nice balancing, certain scenarios can have
6536 * max load less than avg load(as we skip the groups at or below
6537 * its cpu_capacity, while calculating max_load..)
6539 if (busiest->avg_load <= sds->avg_load ||
6540 local->avg_load >= sds->avg_load) {
6542 return fix_small_imbalance(env, sds);
6546 * If there aren't any idle cpus, avoid creating some.
6548 if (busiest->group_type == group_overloaded &&
6549 local->group_type == group_overloaded) {
6550 load_above_capacity = busiest->sum_nr_running *
6552 if (load_above_capacity > busiest->group_capacity)
6553 load_above_capacity -= busiest->group_capacity;
6555 load_above_capacity = ~0UL;
6559 * We're trying to get all the cpus to the average_load, so we don't
6560 * want to push ourselves above the average load, nor do we wish to
6561 * reduce the max loaded cpu below the average load. At the same time,
6562 * we also don't want to reduce the group load below the group capacity
6563 * (so that we can implement power-savings policies etc). Thus we look
6564 * for the minimum possible imbalance.
6566 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
6568 /* How much load to actually move to equalise the imbalance */
6569 env->imbalance = min(
6570 max_pull * busiest->group_capacity,
6571 (sds->avg_load - local->avg_load) * local->group_capacity
6572 ) / SCHED_CAPACITY_SCALE;
6575 * if *imbalance is less than the average load per runnable task
6576 * there is no guarantee that any tasks will be moved so we'll have
6577 * a think about bumping its value to force at least one task to be
6580 if (env->imbalance < busiest->load_per_task)
6581 return fix_small_imbalance(env, sds);
6584 /******* find_busiest_group() helpers end here *********************/
6587 * find_busiest_group - Returns the busiest group within the sched_domain
6588 * if there is an imbalance. If there isn't an imbalance, and
6589 * the user has opted for power-savings, it returns a group whose
6590 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
6591 * such a group exists.
6593 * Also calculates the amount of weighted load which should be moved
6594 * to restore balance.
6596 * @env: The load balancing environment.
6598 * Return: - The busiest group if imbalance exists.
6599 * - If no imbalance and user has opted for power-savings balance,
6600 * return the least loaded group whose CPUs can be
6601 * put to idle by rebalancing its tasks onto our group.
6603 static struct sched_group *find_busiest_group(struct lb_env *env)
6605 struct sg_lb_stats *local, *busiest;
6606 struct sd_lb_stats sds;
6608 init_sd_lb_stats(&sds);
6611 * Compute the various statistics relavent for load balancing at
6614 update_sd_lb_stats(env, &sds);
6615 local = &sds.local_stat;
6616 busiest = &sds.busiest_stat;
6618 /* ASYM feature bypasses nice load balance check */
6619 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
6620 check_asym_packing(env, &sds))
6623 /* There is no busy sibling group to pull tasks from */
6624 if (!sds.busiest || busiest->sum_nr_running == 0)
6627 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
6628 / sds.total_capacity;
6631 * If the busiest group is imbalanced the below checks don't
6632 * work because they assume all things are equal, which typically
6633 * isn't true due to cpus_allowed constraints and the like.
6635 if (busiest->group_type == group_imbalanced)
6638 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
6639 if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
6640 busiest->group_no_capacity)
6644 * If the local group is busier than the selected busiest group
6645 * don't try and pull any tasks.
6647 if (local->avg_load >= busiest->avg_load)
6651 * Don't pull any tasks if this group is already above the domain
6654 if (local->avg_load >= sds.avg_load)
6657 if (env->idle == CPU_IDLE) {
6659 * This cpu is idle. If the busiest group is not overloaded
6660 * and there is no imbalance between this and busiest group
6661 * wrt idle cpus, it is balanced. The imbalance becomes
6662 * significant if the diff is greater than 1 otherwise we
6663 * might end up to just move the imbalance on another group
6665 if ((busiest->group_type != group_overloaded) &&
6666 (local->idle_cpus <= (busiest->idle_cpus + 1)))
6670 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
6671 * imbalance_pct to be conservative.
6673 if (100 * busiest->avg_load <=
6674 env->sd->imbalance_pct * local->avg_load)
6679 /* Looks like there is an imbalance. Compute it */
6680 calculate_imbalance(env, &sds);
6689 * find_busiest_queue - find the busiest runqueue among the cpus in group.
6691 static struct rq *find_busiest_queue(struct lb_env *env,
6692 struct sched_group *group)
6694 struct rq *busiest = NULL, *rq;
6695 unsigned long busiest_load = 0, busiest_capacity = 1;
6698 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6699 unsigned long capacity, wl;
6703 rt = fbq_classify_rq(rq);
6706 * We classify groups/runqueues into three groups:
6707 * - regular: there are !numa tasks
6708 * - remote: there are numa tasks that run on the 'wrong' node
6709 * - all: there is no distinction
6711 * In order to avoid migrating ideally placed numa tasks,
6712 * ignore those when there's better options.
6714 * If we ignore the actual busiest queue to migrate another
6715 * task, the next balance pass can still reduce the busiest
6716 * queue by moving tasks around inside the node.
6718 * If we cannot move enough load due to this classification
6719 * the next pass will adjust the group classification and
6720 * allow migration of more tasks.
6722 * Both cases only affect the total convergence complexity.
6724 if (rt > env->fbq_type)
6727 capacity = capacity_of(i);
6729 wl = weighted_cpuload(i);
6732 * When comparing with imbalance, use weighted_cpuload()
6733 * which is not scaled with the cpu capacity.
6736 if (rq->nr_running == 1 && wl > env->imbalance &&
6737 !check_cpu_capacity(rq, env->sd))
6741 * For the load comparisons with the other cpu's, consider
6742 * the weighted_cpuload() scaled with the cpu capacity, so
6743 * that the load can be moved away from the cpu that is
6744 * potentially running at a lower capacity.
6746 * Thus we're looking for max(wl_i / capacity_i), crosswise
6747 * multiplication to rid ourselves of the division works out
6748 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
6749 * our previous maximum.
6751 if (wl * busiest_capacity > busiest_load * capacity) {
6753 busiest_capacity = capacity;
6762 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6763 * so long as it is large enough.
6765 #define MAX_PINNED_INTERVAL 512
6767 /* Working cpumask for load_balance and load_balance_newidle. */
6768 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6770 static int need_active_balance(struct lb_env *env)
6772 struct sched_domain *sd = env->sd;
6774 if (env->idle == CPU_NEWLY_IDLE) {
6777 * ASYM_PACKING needs to force migrate tasks from busy but
6778 * higher numbered CPUs in order to pack all tasks in the
6779 * lowest numbered CPUs.
6781 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6786 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
6787 * It's worth migrating the task if the src_cpu's capacity is reduced
6788 * because of other sched_class or IRQs if more capacity stays
6789 * available on dst_cpu.
6791 if ((env->idle != CPU_NOT_IDLE) &&
6792 (env->src_rq->cfs.h_nr_running == 1)) {
6793 if ((check_cpu_capacity(env->src_rq, sd)) &&
6794 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
6798 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6801 static int active_load_balance_cpu_stop(void *data);
6803 static int should_we_balance(struct lb_env *env)
6805 struct sched_group *sg = env->sd->groups;
6806 struct cpumask *sg_cpus, *sg_mask;
6807 int cpu, balance_cpu = -1;
6810 * In the newly idle case, we will allow all the cpu's
6811 * to do the newly idle load balance.
6813 if (env->idle == CPU_NEWLY_IDLE)
6816 sg_cpus = sched_group_cpus(sg);
6817 sg_mask = sched_group_mask(sg);
6818 /* Try to find first idle cpu */
6819 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6820 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6827 if (balance_cpu == -1)
6828 balance_cpu = group_balance_cpu(sg);
6831 * First idle cpu or the first cpu(busiest) in this sched group
6832 * is eligible for doing load balancing at this and above domains.
6834 return balance_cpu == env->dst_cpu;
6838 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6839 * tasks if there is an imbalance.
6841 static int load_balance(int this_cpu, struct rq *this_rq,
6842 struct sched_domain *sd, enum cpu_idle_type idle,
6843 int *continue_balancing)
6845 int ld_moved, cur_ld_moved, active_balance = 0;
6846 struct sched_domain *sd_parent = sd->parent;
6847 struct sched_group *group;
6849 unsigned long flags;
6850 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
6852 struct lb_env env = {
6854 .dst_cpu = this_cpu,
6856 .dst_grpmask = sched_group_cpus(sd->groups),
6858 .loop_break = sched_nr_migrate_break,
6861 .tasks = LIST_HEAD_INIT(env.tasks),
6865 * For NEWLY_IDLE load_balancing, we don't need to consider
6866 * other cpus in our group
6868 if (idle == CPU_NEWLY_IDLE)
6869 env.dst_grpmask = NULL;
6871 cpumask_copy(cpus, cpu_active_mask);
6873 schedstat_inc(sd, lb_count[idle]);
6876 if (!should_we_balance(&env)) {
6877 *continue_balancing = 0;
6881 group = find_busiest_group(&env);
6883 schedstat_inc(sd, lb_nobusyg[idle]);
6887 busiest = find_busiest_queue(&env, group);
6889 schedstat_inc(sd, lb_nobusyq[idle]);
6893 BUG_ON(busiest == env.dst_rq);
6895 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6897 env.src_cpu = busiest->cpu;
6898 env.src_rq = busiest;
6901 if (busiest->nr_running > 1) {
6903 * Attempt to move tasks. If find_busiest_group has found
6904 * an imbalance but busiest->nr_running <= 1, the group is
6905 * still unbalanced. ld_moved simply stays zero, so it is
6906 * correctly treated as an imbalance.
6908 env.flags |= LBF_ALL_PINNED;
6909 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6912 raw_spin_lock_irqsave(&busiest->lock, flags);
6915 * cur_ld_moved - load moved in current iteration
6916 * ld_moved - cumulative load moved across iterations
6918 cur_ld_moved = detach_tasks(&env);
6921 * We've detached some tasks from busiest_rq. Every
6922 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
6923 * unlock busiest->lock, and we are able to be sure
6924 * that nobody can manipulate the tasks in parallel.
6925 * See task_rq_lock() family for the details.
6928 raw_spin_unlock(&busiest->lock);
6932 ld_moved += cur_ld_moved;
6935 local_irq_restore(flags);
6937 if (env.flags & LBF_NEED_BREAK) {
6938 env.flags &= ~LBF_NEED_BREAK;
6943 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6944 * us and move them to an alternate dst_cpu in our sched_group
6945 * where they can run. The upper limit on how many times we
6946 * iterate on same src_cpu is dependent on number of cpus in our
6949 * This changes load balance semantics a bit on who can move
6950 * load to a given_cpu. In addition to the given_cpu itself
6951 * (or a ilb_cpu acting on its behalf where given_cpu is
6952 * nohz-idle), we now have balance_cpu in a position to move
6953 * load to given_cpu. In rare situations, this may cause
6954 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6955 * _independently_ and at _same_ time to move some load to
6956 * given_cpu) causing exceess load to be moved to given_cpu.
6957 * This however should not happen so much in practice and
6958 * moreover subsequent load balance cycles should correct the
6959 * excess load moved.
6961 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6963 /* Prevent to re-select dst_cpu via env's cpus */
6964 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6966 env.dst_rq = cpu_rq(env.new_dst_cpu);
6967 env.dst_cpu = env.new_dst_cpu;
6968 env.flags &= ~LBF_DST_PINNED;
6970 env.loop_break = sched_nr_migrate_break;
6973 * Go back to "more_balance" rather than "redo" since we
6974 * need to continue with same src_cpu.
6980 * We failed to reach balance because of affinity.
6983 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
6985 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
6986 *group_imbalance = 1;
6989 /* All tasks on this runqueue were pinned by CPU affinity */
6990 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6991 cpumask_clear_cpu(cpu_of(busiest), cpus);
6992 if (!cpumask_empty(cpus)) {
6994 env.loop_break = sched_nr_migrate_break;
6997 goto out_all_pinned;
7002 schedstat_inc(sd, lb_failed[idle]);
7004 * Increment the failure counter only on periodic balance.
7005 * We do not want newidle balance, which can be very
7006 * frequent, pollute the failure counter causing
7007 * excessive cache_hot migrations and active balances.
7009 if (idle != CPU_NEWLY_IDLE)
7010 sd->nr_balance_failed++;
7012 if (need_active_balance(&env)) {
7013 raw_spin_lock_irqsave(&busiest->lock, flags);
7015 /* don't kick the active_load_balance_cpu_stop,
7016 * if the curr task on busiest cpu can't be
7019 if (!cpumask_test_cpu(this_cpu,
7020 tsk_cpus_allowed(busiest->curr))) {
7021 raw_spin_unlock_irqrestore(&busiest->lock,
7023 env.flags |= LBF_ALL_PINNED;
7024 goto out_one_pinned;
7028 * ->active_balance synchronizes accesses to
7029 * ->active_balance_work. Once set, it's cleared
7030 * only after active load balance is finished.
7032 if (!busiest->active_balance) {
7033 busiest->active_balance = 1;
7034 busiest->push_cpu = this_cpu;
7037 raw_spin_unlock_irqrestore(&busiest->lock, flags);
7039 if (active_balance) {
7040 stop_one_cpu_nowait(cpu_of(busiest),
7041 active_load_balance_cpu_stop, busiest,
7042 &busiest->active_balance_work);
7046 * We've kicked active balancing, reset the failure
7049 sd->nr_balance_failed = sd->cache_nice_tries+1;
7052 sd->nr_balance_failed = 0;
7054 if (likely(!active_balance)) {
7055 /* We were unbalanced, so reset the balancing interval */
7056 sd->balance_interval = sd->min_interval;
7059 * If we've begun active balancing, start to back off. This
7060 * case may not be covered by the all_pinned logic if there
7061 * is only 1 task on the busy runqueue (because we don't call
7064 if (sd->balance_interval < sd->max_interval)
7065 sd->balance_interval *= 2;
7072 * We reach balance although we may have faced some affinity
7073 * constraints. Clear the imbalance flag if it was set.
7076 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7078 if (*group_imbalance)
7079 *group_imbalance = 0;
7084 * We reach balance because all tasks are pinned at this level so
7085 * we can't migrate them. Let the imbalance flag set so parent level
7086 * can try to migrate them.
7088 schedstat_inc(sd, lb_balanced[idle]);
7090 sd->nr_balance_failed = 0;
7093 /* tune up the balancing interval */
7094 if (((env.flags & LBF_ALL_PINNED) &&
7095 sd->balance_interval < MAX_PINNED_INTERVAL) ||
7096 (sd->balance_interval < sd->max_interval))
7097 sd->balance_interval *= 2;
7104 static inline unsigned long
7105 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7107 unsigned long interval = sd->balance_interval;
7110 interval *= sd->busy_factor;
7112 /* scale ms to jiffies */
7113 interval = msecs_to_jiffies(interval);
7114 interval = clamp(interval, 1UL, max_load_balance_interval);
7120 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7122 unsigned long interval, next;
7124 interval = get_sd_balance_interval(sd, cpu_busy);
7125 next = sd->last_balance + interval;
7127 if (time_after(*next_balance, next))
7128 *next_balance = next;
7132 * idle_balance is called by schedule() if this_cpu is about to become
7133 * idle. Attempts to pull tasks from other CPUs.
7135 static int idle_balance(struct rq *this_rq)
7137 unsigned long next_balance = jiffies + HZ;
7138 int this_cpu = this_rq->cpu;
7139 struct sched_domain *sd;
7140 int pulled_task = 0;
7143 idle_enter_fair(this_rq);
7146 * We must set idle_stamp _before_ calling idle_balance(), such that we
7147 * measure the duration of idle_balance() as idle time.
7149 this_rq->idle_stamp = rq_clock(this_rq);
7151 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7152 !this_rq->rd->overload) {
7154 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7156 update_next_balance(sd, 0, &next_balance);
7162 raw_spin_unlock(&this_rq->lock);
7164 update_blocked_averages(this_cpu);
7166 for_each_domain(this_cpu, sd) {
7167 int continue_balancing = 1;
7168 u64 t0, domain_cost;
7170 if (!(sd->flags & SD_LOAD_BALANCE))
7173 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7174 update_next_balance(sd, 0, &next_balance);
7178 if (sd->flags & SD_BALANCE_NEWIDLE) {
7179 t0 = sched_clock_cpu(this_cpu);
7181 pulled_task = load_balance(this_cpu, this_rq,
7183 &continue_balancing);
7185 domain_cost = sched_clock_cpu(this_cpu) - t0;
7186 if (domain_cost > sd->max_newidle_lb_cost)
7187 sd->max_newidle_lb_cost = domain_cost;
7189 curr_cost += domain_cost;
7192 update_next_balance(sd, 0, &next_balance);
7195 * Stop searching for tasks to pull if there are
7196 * now runnable tasks on this rq.
7198 if (pulled_task || this_rq->nr_running > 0)
7203 raw_spin_lock(&this_rq->lock);
7205 if (curr_cost > this_rq->max_idle_balance_cost)
7206 this_rq->max_idle_balance_cost = curr_cost;
7209 * While browsing the domains, we released the rq lock, a task could
7210 * have been enqueued in the meantime. Since we're not going idle,
7211 * pretend we pulled a task.
7213 if (this_rq->cfs.h_nr_running && !pulled_task)
7217 /* Move the next balance forward */
7218 if (time_after(this_rq->next_balance, next_balance))
7219 this_rq->next_balance = next_balance;
7221 /* Is there a task of a high priority class? */
7222 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7226 idle_exit_fair(this_rq);
7227 this_rq->idle_stamp = 0;
7234 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7235 * running tasks off the busiest CPU onto idle CPUs. It requires at
7236 * least 1 task to be running on each physical CPU where possible, and
7237 * avoids physical / logical imbalances.
7239 static int active_load_balance_cpu_stop(void *data)
7241 struct rq *busiest_rq = data;
7242 int busiest_cpu = cpu_of(busiest_rq);
7243 int target_cpu = busiest_rq->push_cpu;
7244 struct rq *target_rq = cpu_rq(target_cpu);
7245 struct sched_domain *sd;
7246 struct task_struct *p = NULL;
7248 raw_spin_lock_irq(&busiest_rq->lock);
7250 /* make sure the requested cpu hasn't gone down in the meantime */
7251 if (unlikely(busiest_cpu != smp_processor_id() ||
7252 !busiest_rq->active_balance))
7255 /* Is there any task to move? */
7256 if (busiest_rq->nr_running <= 1)
7260 * This condition is "impossible", if it occurs
7261 * we need to fix it. Originally reported by
7262 * Bjorn Helgaas on a 128-cpu setup.
7264 BUG_ON(busiest_rq == target_rq);
7266 /* Search for an sd spanning us and the target CPU. */
7268 for_each_domain(target_cpu, sd) {
7269 if ((sd->flags & SD_LOAD_BALANCE) &&
7270 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7275 struct lb_env env = {
7277 .dst_cpu = target_cpu,
7278 .dst_rq = target_rq,
7279 .src_cpu = busiest_rq->cpu,
7280 .src_rq = busiest_rq,
7284 schedstat_inc(sd, alb_count);
7286 p = detach_one_task(&env);
7288 schedstat_inc(sd, alb_pushed);
7290 schedstat_inc(sd, alb_failed);
7294 busiest_rq->active_balance = 0;
7295 raw_spin_unlock(&busiest_rq->lock);
7298 attach_one_task(target_rq, p);
7305 static inline int on_null_domain(struct rq *rq)
7307 return unlikely(!rcu_dereference_sched(rq->sd));
7310 #ifdef CONFIG_NO_HZ_COMMON
7312 * idle load balancing details
7313 * - When one of the busy CPUs notice that there may be an idle rebalancing
7314 * needed, they will kick the idle load balancer, which then does idle
7315 * load balancing for all the idle CPUs.
7318 cpumask_var_t idle_cpus_mask;
7320 unsigned long next_balance; /* in jiffy units */
7321 } nohz ____cacheline_aligned;
7323 static inline int find_new_ilb(void)
7325 int ilb = cpumask_first(nohz.idle_cpus_mask);
7327 if (ilb < nr_cpu_ids && idle_cpu(ilb))
7334 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7335 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7336 * CPU (if there is one).
7338 static void nohz_balancer_kick(void)
7342 nohz.next_balance++;
7344 ilb_cpu = find_new_ilb();
7346 if (ilb_cpu >= nr_cpu_ids)
7349 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7352 * Use smp_send_reschedule() instead of resched_cpu().
7353 * This way we generate a sched IPI on the target cpu which
7354 * is idle. And the softirq performing nohz idle load balance
7355 * will be run before returning from the IPI.
7357 smp_send_reschedule(ilb_cpu);
7361 static inline void nohz_balance_exit_idle(int cpu)
7363 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7365 * Completely isolated CPUs don't ever set, so we must test.
7367 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7368 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7369 atomic_dec(&nohz.nr_cpus);
7371 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7375 static inline void set_cpu_sd_state_busy(void)
7377 struct sched_domain *sd;
7378 int cpu = smp_processor_id();
7381 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7383 if (!sd || !sd->nohz_idle)
7387 atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7392 void set_cpu_sd_state_idle(void)
7394 struct sched_domain *sd;
7395 int cpu = smp_processor_id();
7398 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7400 if (!sd || sd->nohz_idle)
7404 atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7410 * This routine will record that the cpu is going idle with tick stopped.
7411 * This info will be used in performing idle load balancing in the future.
7413 void nohz_balance_enter_idle(int cpu)
7416 * If this cpu is going down, then nothing needs to be done.
7418 if (!cpu_active(cpu))
7421 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
7425 * If we're a completely isolated CPU, we don't play.
7427 if (on_null_domain(cpu_rq(cpu)))
7430 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
7431 atomic_inc(&nohz.nr_cpus);
7432 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7435 static int sched_ilb_notifier(struct notifier_block *nfb,
7436 unsigned long action, void *hcpu)
7438 switch (action & ~CPU_TASKS_FROZEN) {
7440 nohz_balance_exit_idle(smp_processor_id());
7448 static DEFINE_SPINLOCK(balancing);
7451 * Scale the max load_balance interval with the number of CPUs in the system.
7452 * This trades load-balance latency on larger machines for less cross talk.
7454 void update_max_interval(void)
7456 max_load_balance_interval = HZ*num_online_cpus()/10;
7460 * It checks each scheduling domain to see if it is due to be balanced,
7461 * and initiates a balancing operation if so.
7463 * Balancing parameters are set up in init_sched_domains.
7465 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
7467 int continue_balancing = 1;
7469 unsigned long interval;
7470 struct sched_domain *sd;
7471 /* Earliest time when we have to do rebalance again */
7472 unsigned long next_balance = jiffies + 60*HZ;
7473 int update_next_balance = 0;
7474 int need_serialize, need_decay = 0;
7477 update_blocked_averages(cpu);
7480 for_each_domain(cpu, sd) {
7482 * Decay the newidle max times here because this is a regular
7483 * visit to all the domains. Decay ~1% per second.
7485 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
7486 sd->max_newidle_lb_cost =
7487 (sd->max_newidle_lb_cost * 253) / 256;
7488 sd->next_decay_max_lb_cost = jiffies + HZ;
7491 max_cost += sd->max_newidle_lb_cost;
7493 if (!(sd->flags & SD_LOAD_BALANCE))
7497 * Stop the load balance at this level. There is another
7498 * CPU in our sched group which is doing load balancing more
7501 if (!continue_balancing) {
7507 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7509 need_serialize = sd->flags & SD_SERIALIZE;
7510 if (need_serialize) {
7511 if (!spin_trylock(&balancing))
7515 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7516 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
7518 * The LBF_DST_PINNED logic could have changed
7519 * env->dst_cpu, so we can't know our idle
7520 * state even if we migrated tasks. Update it.
7522 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
7524 sd->last_balance = jiffies;
7525 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
7528 spin_unlock(&balancing);
7530 if (time_after(next_balance, sd->last_balance + interval)) {
7531 next_balance = sd->last_balance + interval;
7532 update_next_balance = 1;
7537 * Ensure the rq-wide value also decays but keep it at a
7538 * reasonable floor to avoid funnies with rq->avg_idle.
7540 rq->max_idle_balance_cost =
7541 max((u64)sysctl_sched_migration_cost, max_cost);
7546 * next_balance will be updated only when there is a need.
7547 * When the cpu is attached to null domain for ex, it will not be
7550 if (likely(update_next_balance))
7551 rq->next_balance = next_balance;
7554 #ifdef CONFIG_NO_HZ_COMMON
7556 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
7557 * rebalancing for all the cpus for whom scheduler ticks are stopped.
7559 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
7561 int this_cpu = this_rq->cpu;
7565 if (idle != CPU_IDLE ||
7566 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
7569 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
7570 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
7574 * If this cpu gets work to do, stop the load balancing
7575 * work being done for other cpus. Next load
7576 * balancing owner will pick it up.
7581 rq = cpu_rq(balance_cpu);
7584 * If time for next balance is due,
7587 if (time_after_eq(jiffies, rq->next_balance)) {
7588 raw_spin_lock_irq(&rq->lock);
7589 update_rq_clock(rq);
7590 update_idle_cpu_load(rq);
7591 raw_spin_unlock_irq(&rq->lock);
7592 rebalance_domains(rq, CPU_IDLE);
7595 if (time_after(this_rq->next_balance, rq->next_balance))
7596 this_rq->next_balance = rq->next_balance;
7598 nohz.next_balance = this_rq->next_balance;
7600 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
7604 * Current heuristic for kicking the idle load balancer in the presence
7605 * of an idle cpu in the system.
7606 * - This rq has more than one task.
7607 * - This rq has at least one CFS task and the capacity of the CPU is
7608 * significantly reduced because of RT tasks or IRQs.
7609 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
7610 * multiple busy cpu.
7611 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
7612 * domain span are idle.
7614 static inline bool nohz_kick_needed(struct rq *rq)
7616 unsigned long now = jiffies;
7617 struct sched_domain *sd;
7618 struct sched_group_capacity *sgc;
7619 int nr_busy, cpu = rq->cpu;
7622 if (unlikely(rq->idle_balance))
7626 * We may be recently in ticked or tickless idle mode. At the first
7627 * busy tick after returning from idle, we will update the busy stats.
7629 set_cpu_sd_state_busy();
7630 nohz_balance_exit_idle(cpu);
7633 * None are in tickless mode and hence no need for NOHZ idle load
7636 if (likely(!atomic_read(&nohz.nr_cpus)))
7639 if (time_before(now, nohz.next_balance))
7642 if (rq->nr_running >= 2)
7646 sd = rcu_dereference(per_cpu(sd_busy, cpu));
7648 sgc = sd->groups->sgc;
7649 nr_busy = atomic_read(&sgc->nr_busy_cpus);
7658 sd = rcu_dereference(rq->sd);
7660 if ((rq->cfs.h_nr_running >= 1) &&
7661 check_cpu_capacity(rq, sd)) {
7667 sd = rcu_dereference(per_cpu(sd_asym, cpu));
7668 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
7669 sched_domain_span(sd)) < cpu)) {
7679 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
7683 * run_rebalance_domains is triggered when needed from the scheduler tick.
7684 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
7686 static void run_rebalance_domains(struct softirq_action *h)
7688 struct rq *this_rq = this_rq();
7689 enum cpu_idle_type idle = this_rq->idle_balance ?
7690 CPU_IDLE : CPU_NOT_IDLE;
7693 * If this cpu has a pending nohz_balance_kick, then do the
7694 * balancing on behalf of the other idle cpus whose ticks are
7695 * stopped. Do nohz_idle_balance *before* rebalance_domains to
7696 * give the idle cpus a chance to load balance. Else we may
7697 * load balance only within the local sched_domain hierarchy
7698 * and abort nohz_idle_balance altogether if we pull some load.
7700 nohz_idle_balance(this_rq, idle);
7701 rebalance_domains(this_rq, idle);
7705 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
7707 void trigger_load_balance(struct rq *rq)
7709 /* Don't need to rebalance while attached to NULL domain */
7710 if (unlikely(on_null_domain(rq)))
7713 if (time_after_eq(jiffies, rq->next_balance))
7714 raise_softirq(SCHED_SOFTIRQ);
7715 #ifdef CONFIG_NO_HZ_COMMON
7716 if (nohz_kick_needed(rq))
7717 nohz_balancer_kick();
7721 static void rq_online_fair(struct rq *rq)
7725 update_runtime_enabled(rq);
7728 static void rq_offline_fair(struct rq *rq)
7732 /* Ensure any throttled groups are reachable by pick_next_task */
7733 unthrottle_offline_cfs_rqs(rq);
7736 #endif /* CONFIG_SMP */
7739 * scheduler tick hitting a task of our scheduling class:
7741 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
7743 struct cfs_rq *cfs_rq;
7744 struct sched_entity *se = &curr->se;
7746 for_each_sched_entity(se) {
7747 cfs_rq = cfs_rq_of(se);
7748 entity_tick(cfs_rq, se, queued);
7751 if (numabalancing_enabled)
7752 task_tick_numa(rq, curr);
7756 * called on fork with the child task as argument from the parent's context
7757 * - child not yet on the tasklist
7758 * - preemption disabled
7760 static void task_fork_fair(struct task_struct *p)
7762 struct cfs_rq *cfs_rq;
7763 struct sched_entity *se = &p->se, *curr;
7764 int this_cpu = smp_processor_id();
7765 struct rq *rq = this_rq();
7766 unsigned long flags;
7768 raw_spin_lock_irqsave(&rq->lock, flags);
7770 update_rq_clock(rq);
7772 cfs_rq = task_cfs_rq(current);
7773 curr = cfs_rq->curr;
7776 * Not only the cpu but also the task_group of the parent might have
7777 * been changed after parent->se.parent,cfs_rq were copied to
7778 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
7779 * of child point to valid ones.
7782 __set_task_cpu(p, this_cpu);
7785 update_curr(cfs_rq);
7788 se->vruntime = curr->vruntime;
7789 place_entity(cfs_rq, se, 1);
7791 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
7793 * Upon rescheduling, sched_class::put_prev_task() will place
7794 * 'current' within the tree based on its new key value.
7796 swap(curr->vruntime, se->vruntime);
7800 se->vruntime -= cfs_rq->min_vruntime;
7802 raw_spin_unlock_irqrestore(&rq->lock, flags);
7806 * Priority of the task has changed. Check to see if we preempt
7810 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
7812 if (!task_on_rq_queued(p))
7816 * Reschedule if we are currently running on this runqueue and
7817 * our priority decreased, or if we are not currently running on
7818 * this runqueue and our priority is higher than the current's
7820 if (rq->curr == p) {
7821 if (p->prio > oldprio)
7824 check_preempt_curr(rq, p, 0);
7827 static void switched_from_fair(struct rq *rq, struct task_struct *p)
7829 struct sched_entity *se = &p->se;
7830 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7833 * Ensure the task's vruntime is normalized, so that when it's
7834 * switched back to the fair class the enqueue_entity(.flags=0) will
7835 * do the right thing.
7837 * If it's queued, then the dequeue_entity(.flags=0) will already
7838 * have normalized the vruntime, if it's !queued, then only when
7839 * the task is sleeping will it still have non-normalized vruntime.
7841 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) {
7843 * Fix up our vruntime so that the current sleep doesn't
7844 * cause 'unlimited' sleep bonus.
7846 place_entity(cfs_rq, se, 0);
7847 se->vruntime -= cfs_rq->min_vruntime;
7851 /* Catch up with the cfs_rq and remove our load when we leave */
7852 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq), &se->avg,
7853 se->on_rq * scale_load_down(se->load.weight), cfs_rq->curr == se);
7855 cfs_rq->avg.load_avg =
7856 max_t(long, cfs_rq->avg.load_avg - se->avg.load_avg, 0);
7857 cfs_rq->avg.load_sum =
7858 max_t(s64, cfs_rq->avg.load_sum - se->avg.load_sum, 0);
7859 cfs_rq->avg.util_avg =
7860 max_t(long, cfs_rq->avg.util_avg - se->avg.util_avg, 0);
7861 cfs_rq->avg.util_sum =
7862 max_t(s32, cfs_rq->avg.util_sum - se->avg.util_sum, 0);
7867 * We switched to the sched_fair class.
7869 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7871 #ifdef CONFIG_FAIR_GROUP_SCHED
7872 struct sched_entity *se = &p->se;
7874 * Since the real-depth could have been changed (only FAIR
7875 * class maintain depth value), reset depth properly.
7877 se->depth = se->parent ? se->parent->depth + 1 : 0;
7879 if (!task_on_rq_queued(p))
7883 * We were most likely switched from sched_rt, so
7884 * kick off the schedule if running, otherwise just see
7885 * if we can still preempt the current task.
7890 check_preempt_curr(rq, p, 0);
7893 /* Account for a task changing its policy or group.
7895 * This routine is mostly called to set cfs_rq->curr field when a task
7896 * migrates between groups/classes.
7898 static void set_curr_task_fair(struct rq *rq)
7900 struct sched_entity *se = &rq->curr->se;
7902 for_each_sched_entity(se) {
7903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7905 set_next_entity(cfs_rq, se);
7906 /* ensure bandwidth has been allocated on our new cfs_rq */
7907 account_cfs_rq_runtime(cfs_rq, 0);
7911 void init_cfs_rq(struct cfs_rq *cfs_rq)
7913 cfs_rq->tasks_timeline = RB_ROOT;
7914 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7915 #ifndef CONFIG_64BIT
7916 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7919 atomic_long_set(&cfs_rq->removed_load_avg, 0);
7920 atomic_long_set(&cfs_rq->removed_util_avg, 0);
7924 #ifdef CONFIG_FAIR_GROUP_SCHED
7925 static void task_move_group_fair(struct task_struct *p, int queued)
7927 struct sched_entity *se = &p->se;
7928 struct cfs_rq *cfs_rq;
7931 * If the task was not on the rq at the time of this cgroup movement
7932 * it must have been asleep, sleeping tasks keep their ->vruntime
7933 * absolute on their old rq until wakeup (needed for the fair sleeper
7934 * bonus in place_entity()).
7936 * If it was on the rq, we've just 'preempted' it, which does convert
7937 * ->vruntime to a relative base.
7939 * Make sure both cases convert their relative position when migrating
7940 * to another cgroup's rq. This does somewhat interfere with the
7941 * fair sleeper stuff for the first placement, but who cares.
7944 * When !queued, vruntime of the task has usually NOT been normalized.
7945 * But there are some cases where it has already been normalized:
7947 * - Moving a forked child which is waiting for being woken up by
7948 * wake_up_new_task().
7949 * - Moving a task which has been woken up by try_to_wake_up() and
7950 * waiting for actually being woken up by sched_ttwu_pending().
7952 * To prevent boost or penalty in the new cfs_rq caused by delta
7953 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7955 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING))
7959 se->vruntime -= cfs_rq_of(se)->min_vruntime;
7960 set_task_rq(p, task_cpu(p));
7961 se->depth = se->parent ? se->parent->depth + 1 : 0;
7963 cfs_rq = cfs_rq_of(se);
7964 se->vruntime += cfs_rq->min_vruntime;
7967 /* Virtually synchronize task with its new cfs_rq */
7968 p->se.avg.last_update_time = cfs_rq->avg.last_update_time;
7969 cfs_rq->avg.load_avg += p->se.avg.load_avg;
7970 cfs_rq->avg.load_sum += p->se.avg.load_sum;
7971 cfs_rq->avg.util_avg += p->se.avg.util_avg;
7972 cfs_rq->avg.util_sum += p->se.avg.util_sum;
7977 void free_fair_sched_group(struct task_group *tg)
7981 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7983 for_each_possible_cpu(i) {
7985 kfree(tg->cfs_rq[i]);
7994 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7996 struct cfs_rq *cfs_rq;
7997 struct sched_entity *se;
8000 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8003 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8007 tg->shares = NICE_0_LOAD;
8009 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8011 for_each_possible_cpu(i) {
8012 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8013 GFP_KERNEL, cpu_to_node(i));
8017 se = kzalloc_node(sizeof(struct sched_entity),
8018 GFP_KERNEL, cpu_to_node(i));
8022 init_cfs_rq(cfs_rq);
8023 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8034 void unregister_fair_sched_group(struct task_group *tg, int cpu)
8036 struct rq *rq = cpu_rq(cpu);
8037 unsigned long flags;
8040 * Only empty task groups can be destroyed; so we can speculatively
8041 * check on_list without danger of it being re-added.
8043 if (!tg->cfs_rq[cpu]->on_list)
8046 raw_spin_lock_irqsave(&rq->lock, flags);
8047 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8048 raw_spin_unlock_irqrestore(&rq->lock, flags);
8051 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8052 struct sched_entity *se, int cpu,
8053 struct sched_entity *parent)
8055 struct rq *rq = cpu_rq(cpu);
8059 init_cfs_rq_runtime(cfs_rq);
8061 tg->cfs_rq[cpu] = cfs_rq;
8064 /* se could be NULL for root_task_group */
8069 se->cfs_rq = &rq->cfs;
8072 se->cfs_rq = parent->my_q;
8073 se->depth = parent->depth + 1;
8077 /* guarantee group entities always have weight */
8078 update_load_set(&se->load, NICE_0_LOAD);
8079 se->parent = parent;
8082 static DEFINE_MUTEX(shares_mutex);
8084 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8087 unsigned long flags;
8090 * We can't change the weight of the root cgroup.
8095 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8097 mutex_lock(&shares_mutex);
8098 if (tg->shares == shares)
8101 tg->shares = shares;
8102 for_each_possible_cpu(i) {
8103 struct rq *rq = cpu_rq(i);
8104 struct sched_entity *se;
8107 /* Propagate contribution to hierarchy */
8108 raw_spin_lock_irqsave(&rq->lock, flags);
8110 /* Possible calls to update_curr() need rq clock */
8111 update_rq_clock(rq);
8112 for_each_sched_entity(se)
8113 update_cfs_shares(group_cfs_rq(se));
8114 raw_spin_unlock_irqrestore(&rq->lock, flags);
8118 mutex_unlock(&shares_mutex);
8121 #else /* CONFIG_FAIR_GROUP_SCHED */
8123 void free_fair_sched_group(struct task_group *tg) { }
8125 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8130 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
8132 #endif /* CONFIG_FAIR_GROUP_SCHED */
8135 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
8137 struct sched_entity *se = &task->se;
8138 unsigned int rr_interval = 0;
8141 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
8144 if (rq->cfs.load.weight)
8145 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
8151 * All the scheduling class methods:
8153 const struct sched_class fair_sched_class = {
8154 .next = &idle_sched_class,
8155 .enqueue_task = enqueue_task_fair,
8156 .dequeue_task = dequeue_task_fair,
8157 .yield_task = yield_task_fair,
8158 .yield_to_task = yield_to_task_fair,
8160 .check_preempt_curr = check_preempt_wakeup,
8162 .pick_next_task = pick_next_task_fair,
8163 .put_prev_task = put_prev_task_fair,
8166 .select_task_rq = select_task_rq_fair,
8167 .migrate_task_rq = migrate_task_rq_fair,
8169 .rq_online = rq_online_fair,
8170 .rq_offline = rq_offline_fair,
8172 .task_waking = task_waking_fair,
8175 .set_curr_task = set_curr_task_fair,
8176 .task_tick = task_tick_fair,
8177 .task_fork = task_fork_fair,
8179 .prio_changed = prio_changed_fair,
8180 .switched_from = switched_from_fair,
8181 .switched_to = switched_to_fair,
8183 .get_rr_interval = get_rr_interval_fair,
8185 .update_curr = update_curr_fair,
8187 #ifdef CONFIG_FAIR_GROUP_SCHED
8188 .task_move_group = task_move_group_fair,
8192 #ifdef CONFIG_SCHED_DEBUG
8193 void print_cfs_stats(struct seq_file *m, int cpu)
8195 struct cfs_rq *cfs_rq;
8198 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
8199 print_cfs_rq(m, cpu, cfs_rq);
8203 #ifdef CONFIG_NUMA_BALANCING
8204 void show_numa_stats(struct task_struct *p, struct seq_file *m)
8207 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
8209 for_each_online_node(node) {
8210 if (p->numa_faults) {
8211 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
8212 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
8214 if (p->numa_group) {
8215 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
8216 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
8218 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
8221 #endif /* CONFIG_NUMA_BALANCING */
8222 #endif /* CONFIG_SCHED_DEBUG */
8224 __init void init_sched_fair_class(void)
8227 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8229 #ifdef CONFIG_NO_HZ_COMMON
8230 nohz.next_balance = jiffies;
8231 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8232 cpu_notifier(sched_ilb_notifier, 0);