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/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
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);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
829 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size = 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct *p)
839 unsigned long rss = 0;
840 unsigned long nr_scan_pages;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
848 rss = get_mm_rss(p->mm);
852 rss = round_up(rss, nr_scan_pages);
853 return rss / nr_scan_pages;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct *p)
861 unsigned int scan, floor;
862 unsigned int windows = 1;
864 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
865 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
866 floor = 1000 / windows;
868 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
869 return max_t(unsigned int, floor, scan);
872 static unsigned int task_scan_max(struct task_struct *p)
874 unsigned int smin = task_scan_min(p);
877 /* Watch for min being lower than max due to floor calculations */
878 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
879 return max(smin, smax);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
891 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
893 rq->nr_numa_running += (p->numa_preferred_nid != -1);
894 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
897 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
899 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
900 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
906 spinlock_t lock; /* nr_tasks, tasks */
909 struct list_head task_list;
912 atomic_long_t total_faults;
913 atomic_long_t faults[0];
916 pid_t task_numa_group_id(struct task_struct *p)
918 return p->numa_group ? p->numa_group->gid : 0;
921 static inline int task_faults_idx(int nid, int priv)
923 return 2 * nid + priv;
926 static inline unsigned long task_faults(struct task_struct *p, int nid)
931 return p->numa_faults[task_faults_idx(nid, 0)] +
932 p->numa_faults[task_faults_idx(nid, 1)];
935 static inline unsigned long group_faults(struct task_struct *p, int nid)
940 return atomic_long_read(&p->numa_group->faults[2*nid]) +
941 atomic_long_read(&p->numa_group->faults[2*nid+1]);
945 * These return the fraction of accesses done by a particular task, or
946 * task group, on a particular numa node. The group weight is given a
947 * larger multiplier, in order to group tasks together that are almost
948 * evenly spread out between numa nodes.
950 static inline unsigned long task_weight(struct task_struct *p, int nid)
952 unsigned long total_faults;
957 total_faults = p->total_numa_faults;
962 return 1000 * task_faults(p, nid) / total_faults;
965 static inline unsigned long group_weight(struct task_struct *p, int nid)
967 unsigned long total_faults;
972 total_faults = atomic_long_read(&p->numa_group->total_faults);
977 return 1000 * group_faults(p, nid) / total_faults;
980 static unsigned long weighted_cpuload(const int cpu);
981 static unsigned long source_load(int cpu, int type);
982 static unsigned long target_load(int cpu, int type);
983 static unsigned long power_of(int cpu);
984 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
986 /* Cached statistics for all CPUs within a node */
988 unsigned long nr_running;
991 /* Total compute capacity of CPUs on a node */
994 /* Approximate capacity in terms of runnable tasks on a node */
995 unsigned long capacity;
1000 * XXX borrowed from update_sg_lb_stats
1002 static void update_numa_stats(struct numa_stats *ns, int nid)
1006 memset(ns, 0, sizeof(*ns));
1007 for_each_cpu(cpu, cpumask_of_node(nid)) {
1008 struct rq *rq = cpu_rq(cpu);
1010 ns->nr_running += rq->nr_running;
1011 ns->load += weighted_cpuload(cpu);
1012 ns->power += power_of(cpu);
1015 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1016 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1017 ns->has_capacity = (ns->nr_running < ns->capacity);
1020 struct task_numa_env {
1021 struct task_struct *p;
1023 int src_cpu, src_nid;
1024 int dst_cpu, dst_nid;
1026 struct numa_stats src_stats, dst_stats;
1028 int imbalance_pct, idx;
1030 struct task_struct *best_task;
1035 static void task_numa_assign(struct task_numa_env *env,
1036 struct task_struct *p, long imp)
1039 put_task_struct(env->best_task);
1044 env->best_imp = imp;
1045 env->best_cpu = env->dst_cpu;
1049 * This checks if the overall compute and NUMA accesses of the system would
1050 * be improved if the source tasks was migrated to the target dst_cpu taking
1051 * into account that it might be best if task running on the dst_cpu should
1052 * be exchanged with the source task
1054 static void task_numa_compare(struct task_numa_env *env,
1055 long taskimp, long groupimp)
1057 struct rq *src_rq = cpu_rq(env->src_cpu);
1058 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1059 struct task_struct *cur;
1060 long dst_load, src_load;
1062 long imp = (groupimp > 0) ? groupimp : taskimp;
1065 cur = ACCESS_ONCE(dst_rq->curr);
1066 if (cur->pid == 0) /* idle */
1070 * "imp" is the fault differential for the source task between the
1071 * source and destination node. Calculate the total differential for
1072 * the source task and potential destination task. The more negative
1073 * the value is, the more rmeote accesses that would be expected to
1074 * be incurred if the tasks were swapped.
1077 /* Skip this swap candidate if cannot move to the source cpu */
1078 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1082 * If dst and source tasks are in the same NUMA group, or not
1083 * in any group then look only at task weights.
1085 if (cur->numa_group == env->p->numa_group) {
1086 imp = taskimp + task_weight(cur, env->src_nid) -
1087 task_weight(cur, env->dst_nid);
1089 * Add some hysteresis to prevent swapping the
1090 * tasks within a group over tiny differences.
1092 if (cur->numa_group)
1096 * Compare the group weights. If a task is all by
1097 * itself (not part of a group), use the task weight
1100 if (env->p->numa_group)
1105 if (cur->numa_group)
1106 imp += group_weight(cur, env->src_nid) -
1107 group_weight(cur, env->dst_nid);
1109 imp += task_weight(cur, env->src_nid) -
1110 task_weight(cur, env->dst_nid);
1114 if (imp < env->best_imp)
1118 /* Is there capacity at our destination? */
1119 if (env->src_stats.has_capacity &&
1120 !env->dst_stats.has_capacity)
1126 /* Balance doesn't matter much if we're running a task per cpu */
1127 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1131 * In the overloaded case, try and keep the load balanced.
1134 dst_load = env->dst_stats.load;
1135 src_load = env->src_stats.load;
1137 /* XXX missing power terms */
1138 load = task_h_load(env->p);
1143 load = task_h_load(cur);
1148 /* make src_load the smaller */
1149 if (dst_load < src_load)
1150 swap(dst_load, src_load);
1152 if (src_load * env->imbalance_pct < dst_load * 100)
1156 task_numa_assign(env, cur, imp);
1161 static void task_numa_find_cpu(struct task_numa_env *env,
1162 long taskimp, long groupimp)
1166 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1167 /* Skip this CPU if the source task cannot migrate */
1168 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1172 task_numa_compare(env, taskimp, groupimp);
1176 static int task_numa_migrate(struct task_struct *p)
1178 struct task_numa_env env = {
1181 .src_cpu = task_cpu(p),
1182 .src_nid = task_node(p),
1184 .imbalance_pct = 112,
1190 struct sched_domain *sd;
1191 unsigned long taskweight, groupweight;
1193 long taskimp, groupimp;
1196 * Pick the lowest SD_NUMA domain, as that would have the smallest
1197 * imbalance and would be the first to start moving tasks about.
1199 * And we want to avoid any moving of tasks about, as that would create
1200 * random movement of tasks -- counter the numa conditions we're trying
1204 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1205 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1208 taskweight = task_weight(p, env.src_nid);
1209 groupweight = group_weight(p, env.src_nid);
1210 update_numa_stats(&env.src_stats, env.src_nid);
1211 env.dst_nid = p->numa_preferred_nid;
1212 taskimp = task_weight(p, env.dst_nid) - taskweight;
1213 groupimp = group_weight(p, env.dst_nid) - groupweight;
1214 update_numa_stats(&env.dst_stats, env.dst_nid);
1216 /* If the preferred nid has capacity, try to use it. */
1217 if (env.dst_stats.has_capacity)
1218 task_numa_find_cpu(&env, taskimp, groupimp);
1220 /* No space available on the preferred nid. Look elsewhere. */
1221 if (env.best_cpu == -1) {
1222 for_each_online_node(nid) {
1223 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1226 /* Only consider nodes where both task and groups benefit */
1227 taskimp = task_weight(p, nid) - taskweight;
1228 groupimp = group_weight(p, nid) - groupweight;
1229 if (taskimp < 0 && groupimp < 0)
1233 update_numa_stats(&env.dst_stats, env.dst_nid);
1234 task_numa_find_cpu(&env, taskimp, groupimp);
1238 /* No better CPU than the current one was found. */
1239 if (env.best_cpu == -1)
1242 sched_setnuma(p, env.dst_nid);
1244 if (env.best_task == NULL) {
1245 int ret = migrate_task_to(p, env.best_cpu);
1249 ret = migrate_swap(p, env.best_task);
1250 put_task_struct(env.best_task);
1254 /* Attempt to migrate a task to a CPU on the preferred node. */
1255 static void numa_migrate_preferred(struct task_struct *p)
1257 /* Success if task is already running on preferred CPU */
1258 p->numa_migrate_retry = 0;
1259 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1261 * If migration is temporarily disabled due to a task migration
1262 * then re-enable it now as the task is running on its
1263 * preferred node and memory should migrate locally
1265 if (!p->numa_migrate_seq)
1266 p->numa_migrate_seq++;
1270 /* This task has no NUMA fault statistics yet */
1271 if (unlikely(p->numa_preferred_nid == -1))
1274 /* Otherwise, try migrate to a CPU on the preferred node */
1275 if (task_numa_migrate(p) != 0)
1276 p->numa_migrate_retry = jiffies + HZ*5;
1279 static void task_numa_placement(struct task_struct *p)
1281 int seq, nid, max_nid = -1, max_group_nid = -1;
1282 unsigned long max_faults = 0, max_group_faults = 0;
1283 spinlock_t *group_lock = NULL;
1285 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1286 if (p->numa_scan_seq == seq)
1288 p->numa_scan_seq = seq;
1289 p->numa_migrate_seq++;
1290 p->numa_scan_period_max = task_scan_max(p);
1292 /* If the task is part of a group prevent parallel updates to group stats */
1293 if (p->numa_group) {
1294 group_lock = &p->numa_group->lock;
1295 spin_lock(group_lock);
1298 /* Find the node with the highest number of faults */
1299 for_each_online_node(nid) {
1300 unsigned long faults = 0, group_faults = 0;
1303 for (priv = 0; priv < 2; priv++) {
1306 i = task_faults_idx(nid, priv);
1307 diff = -p->numa_faults[i];
1309 /* Decay existing window, copy faults since last scan */
1310 p->numa_faults[i] >>= 1;
1311 p->numa_faults[i] += p->numa_faults_buffer[i];
1312 p->numa_faults_buffer[i] = 0;
1314 faults += p->numa_faults[i];
1315 diff += p->numa_faults[i];
1316 p->total_numa_faults += diff;
1317 if (p->numa_group) {
1318 /* safe because we can only change our own group */
1319 atomic_long_add(diff, &p->numa_group->faults[i]);
1320 atomic_long_add(diff, &p->numa_group->total_faults);
1321 group_faults += atomic_long_read(&p->numa_group->faults[i]);
1325 if (faults > max_faults) {
1326 max_faults = faults;
1330 if (group_faults > max_group_faults) {
1331 max_group_faults = group_faults;
1332 max_group_nid = nid;
1336 if (p->numa_group) {
1338 * If the preferred task and group nids are different,
1339 * iterate over the nodes again to find the best place.
1341 if (max_nid != max_group_nid) {
1342 unsigned long weight, max_weight = 0;
1344 for_each_online_node(nid) {
1345 weight = task_weight(p, nid) + group_weight(p, nid);
1346 if (weight > max_weight) {
1347 max_weight = weight;
1353 spin_unlock(group_lock);
1356 /* Preferred node as the node with the most faults */
1357 if (max_faults && max_nid != p->numa_preferred_nid) {
1358 /* Update the preferred nid and migrate task if possible */
1359 sched_setnuma(p, max_nid);
1360 numa_migrate_preferred(p);
1364 static inline int get_numa_group(struct numa_group *grp)
1366 return atomic_inc_not_zero(&grp->refcount);
1369 static inline void put_numa_group(struct numa_group *grp)
1371 if (atomic_dec_and_test(&grp->refcount))
1372 kfree_rcu(grp, rcu);
1375 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1381 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1384 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1387 struct numa_group *grp, *my_grp;
1388 struct task_struct *tsk;
1390 int cpu = cpupid_to_cpu(cpupid);
1393 if (unlikely(!p->numa_group)) {
1394 unsigned int size = sizeof(struct numa_group) +
1395 2*nr_node_ids*sizeof(atomic_long_t);
1397 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1401 atomic_set(&grp->refcount, 1);
1402 spin_lock_init(&grp->lock);
1403 INIT_LIST_HEAD(&grp->task_list);
1406 for (i = 0; i < 2*nr_node_ids; i++)
1407 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1409 atomic_long_set(&grp->total_faults, p->total_numa_faults);
1411 list_add(&p->numa_entry, &grp->task_list);
1413 rcu_assign_pointer(p->numa_group, grp);
1417 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1419 if (!cpupid_match_pid(tsk, cpupid))
1422 grp = rcu_dereference(tsk->numa_group);
1426 my_grp = p->numa_group;
1431 * Only join the other group if its bigger; if we're the bigger group,
1432 * the other task will join us.
1434 if (my_grp->nr_tasks > grp->nr_tasks)
1438 * Tie-break on the grp address.
1440 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1443 /* Always join threads in the same process. */
1444 if (tsk->mm == current->mm)
1447 /* Simple filter to avoid false positives due to PID collisions */
1448 if (flags & TNF_SHARED)
1451 /* Update priv based on whether false sharing was detected */
1454 if (join && !get_numa_group(grp))
1463 for (i = 0; i < 2*nr_node_ids; i++) {
1464 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1465 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1467 atomic_long_sub(p->total_numa_faults, &my_grp->total_faults);
1468 atomic_long_add(p->total_numa_faults, &grp->total_faults);
1470 double_lock(&my_grp->lock, &grp->lock);
1472 list_move(&p->numa_entry, &grp->task_list);
1476 spin_unlock(&my_grp->lock);
1477 spin_unlock(&grp->lock);
1479 rcu_assign_pointer(p->numa_group, grp);
1481 put_numa_group(my_grp);
1484 void task_numa_free(struct task_struct *p)
1486 struct numa_group *grp = p->numa_group;
1488 void *numa_faults = p->numa_faults;
1491 for (i = 0; i < 2*nr_node_ids; i++)
1492 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1494 atomic_long_sub(p->total_numa_faults, &grp->total_faults);
1496 spin_lock(&grp->lock);
1497 list_del(&p->numa_entry);
1499 spin_unlock(&grp->lock);
1500 rcu_assign_pointer(p->numa_group, NULL);
1501 put_numa_group(grp);
1504 p->numa_faults = NULL;
1505 p->numa_faults_buffer = NULL;
1510 * Got a PROT_NONE fault for a page on @node.
1512 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1514 struct task_struct *p = current;
1515 bool migrated = flags & TNF_MIGRATED;
1518 if (!numabalancing_enabled)
1521 /* for example, ksmd faulting in a user's mm */
1525 /* Do not worry about placement if exiting */
1526 if (p->state == TASK_DEAD)
1529 /* Allocate buffer to track faults on a per-node basis */
1530 if (unlikely(!p->numa_faults)) {
1531 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1533 /* numa_faults and numa_faults_buffer share the allocation */
1534 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1535 if (!p->numa_faults)
1538 BUG_ON(p->numa_faults_buffer);
1539 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1540 p->total_numa_faults = 0;
1544 * First accesses are treated as private, otherwise consider accesses
1545 * to be private if the accessing pid has not changed
1547 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1550 priv = cpupid_match_pid(p, last_cpupid);
1551 if (!priv && !(flags & TNF_NO_GROUP))
1552 task_numa_group(p, last_cpupid, flags, &priv);
1556 * If pages are properly placed (did not migrate) then scan slower.
1557 * This is reset periodically in case of phase changes
1560 /* Initialise if necessary */
1561 if (!p->numa_scan_period_max)
1562 p->numa_scan_period_max = task_scan_max(p);
1564 p->numa_scan_period = min(p->numa_scan_period_max,
1565 p->numa_scan_period + 10);
1568 task_numa_placement(p);
1570 /* Retry task to preferred node migration if it previously failed */
1571 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1572 numa_migrate_preferred(p);
1575 p->numa_pages_migrated += pages;
1577 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1580 static void reset_ptenuma_scan(struct task_struct *p)
1582 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1583 p->mm->numa_scan_offset = 0;
1587 * The expensive part of numa migration is done from task_work context.
1588 * Triggered from task_tick_numa().
1590 void task_numa_work(struct callback_head *work)
1592 unsigned long migrate, next_scan, now = jiffies;
1593 struct task_struct *p = current;
1594 struct mm_struct *mm = p->mm;
1595 struct vm_area_struct *vma;
1596 unsigned long start, end;
1597 unsigned long nr_pte_updates = 0;
1600 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1602 work->next = work; /* protect against double add */
1604 * Who cares about NUMA placement when they're dying.
1606 * NOTE: make sure not to dereference p->mm before this check,
1607 * exit_task_work() happens _after_ exit_mm() so we could be called
1608 * without p->mm even though we still had it when we enqueued this
1611 if (p->flags & PF_EXITING)
1614 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1615 mm->numa_next_scan = now +
1616 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1617 mm->numa_next_reset = now +
1618 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1622 * Reset the scan period if enough time has gone by. Objective is that
1623 * scanning will be reduced if pages are properly placed. As tasks
1624 * can enter different phases this needs to be re-examined. Lacking
1625 * proper tracking of reference behaviour, this blunt hammer is used.
1627 migrate = mm->numa_next_reset;
1628 if (time_after(now, migrate)) {
1629 p->numa_scan_period = task_scan_min(p);
1630 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1631 xchg(&mm->numa_next_reset, next_scan);
1635 * Enforce maximal scan/migration frequency..
1637 migrate = mm->numa_next_scan;
1638 if (time_before(now, migrate))
1641 if (p->numa_scan_period == 0) {
1642 p->numa_scan_period_max = task_scan_max(p);
1643 p->numa_scan_period = task_scan_min(p);
1646 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1647 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1651 * Delay this task enough that another task of this mm will likely win
1652 * the next time around.
1654 p->node_stamp += 2 * TICK_NSEC;
1656 start = mm->numa_scan_offset;
1657 pages = sysctl_numa_balancing_scan_size;
1658 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1662 down_read(&mm->mmap_sem);
1663 vma = find_vma(mm, start);
1665 reset_ptenuma_scan(p);
1669 for (; vma; vma = vma->vm_next) {
1670 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1674 * Shared library pages mapped by multiple processes are not
1675 * migrated as it is expected they are cache replicated. Avoid
1676 * hinting faults in read-only file-backed mappings or the vdso
1677 * as migrating the pages will be of marginal benefit.
1680 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1684 start = max(start, vma->vm_start);
1685 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1686 end = min(end, vma->vm_end);
1687 nr_pte_updates += change_prot_numa(vma, start, end);
1690 * Scan sysctl_numa_balancing_scan_size but ensure that
1691 * at least one PTE is updated so that unused virtual
1692 * address space is quickly skipped.
1695 pages -= (end - start) >> PAGE_SHIFT;
1700 } while (end != vma->vm_end);
1705 * If the whole process was scanned without updates then no NUMA
1706 * hinting faults are being recorded and scan rate should be lower.
1708 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1709 p->numa_scan_period = min(p->numa_scan_period_max,
1710 p->numa_scan_period << 1);
1712 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1713 mm->numa_next_scan = next_scan;
1717 * It is possible to reach the end of the VMA list but the last few
1718 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1719 * would find the !migratable VMA on the next scan but not reset the
1720 * scanner to the start so check it now.
1723 mm->numa_scan_offset = start;
1725 reset_ptenuma_scan(p);
1726 up_read(&mm->mmap_sem);
1730 * Drive the periodic memory faults..
1732 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1734 struct callback_head *work = &curr->numa_work;
1738 * We don't care about NUMA placement if we don't have memory.
1740 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1744 * Using runtime rather than walltime has the dual advantage that
1745 * we (mostly) drive the selection from busy threads and that the
1746 * task needs to have done some actual work before we bother with
1749 now = curr->se.sum_exec_runtime;
1750 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1752 if (now - curr->node_stamp > period) {
1753 if (!curr->node_stamp)
1754 curr->numa_scan_period = task_scan_min(curr);
1755 curr->node_stamp += period;
1757 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1758 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1759 task_work_add(curr, work, true);
1764 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1768 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1772 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1775 #endif /* CONFIG_NUMA_BALANCING */
1778 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1780 update_load_add(&cfs_rq->load, se->load.weight);
1781 if (!parent_entity(se))
1782 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1784 if (entity_is_task(se)) {
1785 struct rq *rq = rq_of(cfs_rq);
1787 account_numa_enqueue(rq, task_of(se));
1788 list_add(&se->group_node, &rq->cfs_tasks);
1791 cfs_rq->nr_running++;
1795 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1797 update_load_sub(&cfs_rq->load, se->load.weight);
1798 if (!parent_entity(se))
1799 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1800 if (entity_is_task(se)) {
1801 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1802 list_del_init(&se->group_node);
1804 cfs_rq->nr_running--;
1807 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1814 * Use this CPU's actual weight instead of the last load_contribution
1815 * to gain a more accurate current total weight. See
1816 * update_cfs_rq_load_contribution().
1818 tg_weight = atomic_long_read(&tg->load_avg);
1819 tg_weight -= cfs_rq->tg_load_contrib;
1820 tg_weight += cfs_rq->load.weight;
1825 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1827 long tg_weight, load, shares;
1829 tg_weight = calc_tg_weight(tg, cfs_rq);
1830 load = cfs_rq->load.weight;
1832 shares = (tg->shares * load);
1834 shares /= tg_weight;
1836 if (shares < MIN_SHARES)
1837 shares = MIN_SHARES;
1838 if (shares > tg->shares)
1839 shares = tg->shares;
1843 # else /* CONFIG_SMP */
1844 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1848 # endif /* CONFIG_SMP */
1849 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1850 unsigned long weight)
1853 /* commit outstanding execution time */
1854 if (cfs_rq->curr == se)
1855 update_curr(cfs_rq);
1856 account_entity_dequeue(cfs_rq, se);
1859 update_load_set(&se->load, weight);
1862 account_entity_enqueue(cfs_rq, se);
1865 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1867 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1869 struct task_group *tg;
1870 struct sched_entity *se;
1874 se = tg->se[cpu_of(rq_of(cfs_rq))];
1875 if (!se || throttled_hierarchy(cfs_rq))
1878 if (likely(se->load.weight == tg->shares))
1881 shares = calc_cfs_shares(cfs_rq, tg);
1883 reweight_entity(cfs_rq_of(se), se, shares);
1885 #else /* CONFIG_FAIR_GROUP_SCHED */
1886 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1889 #endif /* CONFIG_FAIR_GROUP_SCHED */
1893 * We choose a half-life close to 1 scheduling period.
1894 * Note: The tables below are dependent on this value.
1896 #define LOAD_AVG_PERIOD 32
1897 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1898 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1900 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1901 static const u32 runnable_avg_yN_inv[] = {
1902 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1903 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1904 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1905 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1906 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1907 0x85aac367, 0x82cd8698,
1911 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1912 * over-estimates when re-combining.
1914 static const u32 runnable_avg_yN_sum[] = {
1915 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1916 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1917 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1922 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1924 static __always_inline u64 decay_load(u64 val, u64 n)
1926 unsigned int local_n;
1930 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1933 /* after bounds checking we can collapse to 32-bit */
1937 * As y^PERIOD = 1/2, we can combine
1938 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1939 * With a look-up table which covers k^n (n<PERIOD)
1941 * To achieve constant time decay_load.
1943 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1944 val >>= local_n / LOAD_AVG_PERIOD;
1945 local_n %= LOAD_AVG_PERIOD;
1948 val *= runnable_avg_yN_inv[local_n];
1949 /* We don't use SRR here since we always want to round down. */
1954 * For updates fully spanning n periods, the contribution to runnable
1955 * average will be: \Sum 1024*y^n
1957 * We can compute this reasonably efficiently by combining:
1958 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1960 static u32 __compute_runnable_contrib(u64 n)
1964 if (likely(n <= LOAD_AVG_PERIOD))
1965 return runnable_avg_yN_sum[n];
1966 else if (unlikely(n >= LOAD_AVG_MAX_N))
1967 return LOAD_AVG_MAX;
1969 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1971 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1972 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1974 n -= LOAD_AVG_PERIOD;
1975 } while (n > LOAD_AVG_PERIOD);
1977 contrib = decay_load(contrib, n);
1978 return contrib + runnable_avg_yN_sum[n];
1982 * We can represent the historical contribution to runnable average as the
1983 * coefficients of a geometric series. To do this we sub-divide our runnable
1984 * history into segments of approximately 1ms (1024us); label the segment that
1985 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1987 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1989 * (now) (~1ms ago) (~2ms ago)
1991 * Let u_i denote the fraction of p_i that the entity was runnable.
1993 * We then designate the fractions u_i as our co-efficients, yielding the
1994 * following representation of historical load:
1995 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1997 * We choose y based on the with of a reasonably scheduling period, fixing:
2000 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2001 * approximately half as much as the contribution to load within the last ms
2004 * When a period "rolls over" and we have new u_0`, multiplying the previous
2005 * sum again by y is sufficient to update:
2006 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2007 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2009 static __always_inline int __update_entity_runnable_avg(u64 now,
2010 struct sched_avg *sa,
2014 u32 runnable_contrib;
2015 int delta_w, decayed = 0;
2017 delta = now - sa->last_runnable_update;
2019 * This should only happen when time goes backwards, which it
2020 * unfortunately does during sched clock init when we swap over to TSC.
2022 if ((s64)delta < 0) {
2023 sa->last_runnable_update = now;
2028 * Use 1024ns as the unit of measurement since it's a reasonable
2029 * approximation of 1us and fast to compute.
2034 sa->last_runnable_update = now;
2036 /* delta_w is the amount already accumulated against our next period */
2037 delta_w = sa->runnable_avg_period % 1024;
2038 if (delta + delta_w >= 1024) {
2039 /* period roll-over */
2043 * Now that we know we're crossing a period boundary, figure
2044 * out how much from delta we need to complete the current
2045 * period and accrue it.
2047 delta_w = 1024 - delta_w;
2049 sa->runnable_avg_sum += delta_w;
2050 sa->runnable_avg_period += delta_w;
2054 /* Figure out how many additional periods this update spans */
2055 periods = delta / 1024;
2058 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2060 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2063 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2064 runnable_contrib = __compute_runnable_contrib(periods);
2066 sa->runnable_avg_sum += runnable_contrib;
2067 sa->runnable_avg_period += runnable_contrib;
2070 /* Remainder of delta accrued against u_0` */
2072 sa->runnable_avg_sum += delta;
2073 sa->runnable_avg_period += delta;
2078 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2079 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2082 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2084 decays -= se->avg.decay_count;
2088 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2089 se->avg.decay_count = 0;
2094 #ifdef CONFIG_FAIR_GROUP_SCHED
2095 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2098 struct task_group *tg = cfs_rq->tg;
2101 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2102 tg_contrib -= cfs_rq->tg_load_contrib;
2104 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2105 atomic_long_add(tg_contrib, &tg->load_avg);
2106 cfs_rq->tg_load_contrib += tg_contrib;
2111 * Aggregate cfs_rq runnable averages into an equivalent task_group
2112 * representation for computing load contributions.
2114 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2115 struct cfs_rq *cfs_rq)
2117 struct task_group *tg = cfs_rq->tg;
2120 /* The fraction of a cpu used by this cfs_rq */
2121 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2122 sa->runnable_avg_period + 1);
2123 contrib -= cfs_rq->tg_runnable_contrib;
2125 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2126 atomic_add(contrib, &tg->runnable_avg);
2127 cfs_rq->tg_runnable_contrib += contrib;
2131 static inline void __update_group_entity_contrib(struct sched_entity *se)
2133 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2134 struct task_group *tg = cfs_rq->tg;
2139 contrib = cfs_rq->tg_load_contrib * tg->shares;
2140 se->avg.load_avg_contrib = div_u64(contrib,
2141 atomic_long_read(&tg->load_avg) + 1);
2144 * For group entities we need to compute a correction term in the case
2145 * that they are consuming <1 cpu so that we would contribute the same
2146 * load as a task of equal weight.
2148 * Explicitly co-ordinating this measurement would be expensive, but
2149 * fortunately the sum of each cpus contribution forms a usable
2150 * lower-bound on the true value.
2152 * Consider the aggregate of 2 contributions. Either they are disjoint
2153 * (and the sum represents true value) or they are disjoint and we are
2154 * understating by the aggregate of their overlap.
2156 * Extending this to N cpus, for a given overlap, the maximum amount we
2157 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2158 * cpus that overlap for this interval and w_i is the interval width.
2160 * On a small machine; the first term is well-bounded which bounds the
2161 * total error since w_i is a subset of the period. Whereas on a
2162 * larger machine, while this first term can be larger, if w_i is the
2163 * of consequential size guaranteed to see n_i*w_i quickly converge to
2164 * our upper bound of 1-cpu.
2166 runnable_avg = atomic_read(&tg->runnable_avg);
2167 if (runnable_avg < NICE_0_LOAD) {
2168 se->avg.load_avg_contrib *= runnable_avg;
2169 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2173 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2174 int force_update) {}
2175 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2176 struct cfs_rq *cfs_rq) {}
2177 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2180 static inline void __update_task_entity_contrib(struct sched_entity *se)
2184 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2185 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2186 contrib /= (se->avg.runnable_avg_period + 1);
2187 se->avg.load_avg_contrib = scale_load(contrib);
2190 /* Compute the current contribution to load_avg by se, return any delta */
2191 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2193 long old_contrib = se->avg.load_avg_contrib;
2195 if (entity_is_task(se)) {
2196 __update_task_entity_contrib(se);
2198 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2199 __update_group_entity_contrib(se);
2202 return se->avg.load_avg_contrib - old_contrib;
2205 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2208 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2209 cfs_rq->blocked_load_avg -= load_contrib;
2211 cfs_rq->blocked_load_avg = 0;
2214 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2216 /* Update a sched_entity's runnable average */
2217 static inline void update_entity_load_avg(struct sched_entity *se,
2220 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2225 * For a group entity we need to use their owned cfs_rq_clock_task() in
2226 * case they are the parent of a throttled hierarchy.
2228 if (entity_is_task(se))
2229 now = cfs_rq_clock_task(cfs_rq);
2231 now = cfs_rq_clock_task(group_cfs_rq(se));
2233 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2236 contrib_delta = __update_entity_load_avg_contrib(se);
2242 cfs_rq->runnable_load_avg += contrib_delta;
2244 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2248 * Decay the load contributed by all blocked children and account this so that
2249 * their contribution may appropriately discounted when they wake up.
2251 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2253 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2256 decays = now - cfs_rq->last_decay;
2257 if (!decays && !force_update)
2260 if (atomic_long_read(&cfs_rq->removed_load)) {
2261 unsigned long removed_load;
2262 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2263 subtract_blocked_load_contrib(cfs_rq, removed_load);
2267 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2269 atomic64_add(decays, &cfs_rq->decay_counter);
2270 cfs_rq->last_decay = now;
2273 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2276 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2278 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2279 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2282 /* Add the load generated by se into cfs_rq's child load-average */
2283 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2284 struct sched_entity *se,
2288 * We track migrations using entity decay_count <= 0, on a wake-up
2289 * migration we use a negative decay count to track the remote decays
2290 * accumulated while sleeping.
2292 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2293 * are seen by enqueue_entity_load_avg() as a migration with an already
2294 * constructed load_avg_contrib.
2296 if (unlikely(se->avg.decay_count <= 0)) {
2297 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2298 if (se->avg.decay_count) {
2300 * In a wake-up migration we have to approximate the
2301 * time sleeping. This is because we can't synchronize
2302 * clock_task between the two cpus, and it is not
2303 * guaranteed to be read-safe. Instead, we can
2304 * approximate this using our carried decays, which are
2305 * explicitly atomically readable.
2307 se->avg.last_runnable_update -= (-se->avg.decay_count)
2309 update_entity_load_avg(se, 0);
2310 /* Indicate that we're now synchronized and on-rq */
2311 se->avg.decay_count = 0;
2316 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2317 * would have made count negative); we must be careful to avoid
2318 * double-accounting blocked time after synchronizing decays.
2320 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2324 /* migrated tasks did not contribute to our blocked load */
2326 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2327 update_entity_load_avg(se, 0);
2330 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2331 /* we force update consideration on load-balancer moves */
2332 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2336 * Remove se's load from this cfs_rq child load-average, if the entity is
2337 * transitioning to a blocked state we track its projected decay using
2340 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2341 struct sched_entity *se,
2344 update_entity_load_avg(se, 1);
2345 /* we force update consideration on load-balancer moves */
2346 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2348 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2350 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2351 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2352 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2356 * Update the rq's load with the elapsed running time before entering
2357 * idle. if the last scheduled task is not a CFS task, idle_enter will
2358 * be the only way to update the runnable statistic.
2360 void idle_enter_fair(struct rq *this_rq)
2362 update_rq_runnable_avg(this_rq, 1);
2366 * Update the rq's load with the elapsed idle time before a task is
2367 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2368 * be the only way to update the runnable statistic.
2370 void idle_exit_fair(struct rq *this_rq)
2372 update_rq_runnable_avg(this_rq, 0);
2376 static inline void update_entity_load_avg(struct sched_entity *se,
2377 int update_cfs_rq) {}
2378 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2379 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2380 struct sched_entity *se,
2382 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2383 struct sched_entity *se,
2385 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2386 int force_update) {}
2389 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2391 #ifdef CONFIG_SCHEDSTATS
2392 struct task_struct *tsk = NULL;
2394 if (entity_is_task(se))
2397 if (se->statistics.sleep_start) {
2398 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2403 if (unlikely(delta > se->statistics.sleep_max))
2404 se->statistics.sleep_max = delta;
2406 se->statistics.sleep_start = 0;
2407 se->statistics.sum_sleep_runtime += delta;
2410 account_scheduler_latency(tsk, delta >> 10, 1);
2411 trace_sched_stat_sleep(tsk, delta);
2414 if (se->statistics.block_start) {
2415 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2420 if (unlikely(delta > se->statistics.block_max))
2421 se->statistics.block_max = delta;
2423 se->statistics.block_start = 0;
2424 se->statistics.sum_sleep_runtime += delta;
2427 if (tsk->in_iowait) {
2428 se->statistics.iowait_sum += delta;
2429 se->statistics.iowait_count++;
2430 trace_sched_stat_iowait(tsk, delta);
2433 trace_sched_stat_blocked(tsk, delta);
2436 * Blocking time is in units of nanosecs, so shift by
2437 * 20 to get a milliseconds-range estimation of the
2438 * amount of time that the task spent sleeping:
2440 if (unlikely(prof_on == SLEEP_PROFILING)) {
2441 profile_hits(SLEEP_PROFILING,
2442 (void *)get_wchan(tsk),
2445 account_scheduler_latency(tsk, delta >> 10, 0);
2451 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2453 #ifdef CONFIG_SCHED_DEBUG
2454 s64 d = se->vruntime - cfs_rq->min_vruntime;
2459 if (d > 3*sysctl_sched_latency)
2460 schedstat_inc(cfs_rq, nr_spread_over);
2465 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2467 u64 vruntime = cfs_rq->min_vruntime;
2470 * The 'current' period is already promised to the current tasks,
2471 * however the extra weight of the new task will slow them down a
2472 * little, place the new task so that it fits in the slot that
2473 * stays open at the end.
2475 if (initial && sched_feat(START_DEBIT))
2476 vruntime += sched_vslice(cfs_rq, se);
2478 /* sleeps up to a single latency don't count. */
2480 unsigned long thresh = sysctl_sched_latency;
2483 * Halve their sleep time's effect, to allow
2484 * for a gentler effect of sleepers:
2486 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2492 /* ensure we never gain time by being placed backwards. */
2493 se->vruntime = max_vruntime(se->vruntime, vruntime);
2496 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2499 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2502 * Update the normalized vruntime before updating min_vruntime
2503 * through calling update_curr().
2505 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2506 se->vruntime += cfs_rq->min_vruntime;
2509 * Update run-time statistics of the 'current'.
2511 update_curr(cfs_rq);
2512 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2513 account_entity_enqueue(cfs_rq, se);
2514 update_cfs_shares(cfs_rq);
2516 if (flags & ENQUEUE_WAKEUP) {
2517 place_entity(cfs_rq, se, 0);
2518 enqueue_sleeper(cfs_rq, se);
2521 update_stats_enqueue(cfs_rq, se);
2522 check_spread(cfs_rq, se);
2523 if (se != cfs_rq->curr)
2524 __enqueue_entity(cfs_rq, se);
2527 if (cfs_rq->nr_running == 1) {
2528 list_add_leaf_cfs_rq(cfs_rq);
2529 check_enqueue_throttle(cfs_rq);
2533 static void __clear_buddies_last(struct sched_entity *se)
2535 for_each_sched_entity(se) {
2536 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2537 if (cfs_rq->last == se)
2538 cfs_rq->last = NULL;
2544 static void __clear_buddies_next(struct sched_entity *se)
2546 for_each_sched_entity(se) {
2547 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2548 if (cfs_rq->next == se)
2549 cfs_rq->next = NULL;
2555 static void __clear_buddies_skip(struct sched_entity *se)
2557 for_each_sched_entity(se) {
2558 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2559 if (cfs_rq->skip == se)
2560 cfs_rq->skip = NULL;
2566 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2568 if (cfs_rq->last == se)
2569 __clear_buddies_last(se);
2571 if (cfs_rq->next == se)
2572 __clear_buddies_next(se);
2574 if (cfs_rq->skip == se)
2575 __clear_buddies_skip(se);
2578 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2581 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2584 * Update run-time statistics of the 'current'.
2586 update_curr(cfs_rq);
2587 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2589 update_stats_dequeue(cfs_rq, se);
2590 if (flags & DEQUEUE_SLEEP) {
2591 #ifdef CONFIG_SCHEDSTATS
2592 if (entity_is_task(se)) {
2593 struct task_struct *tsk = task_of(se);
2595 if (tsk->state & TASK_INTERRUPTIBLE)
2596 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2597 if (tsk->state & TASK_UNINTERRUPTIBLE)
2598 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2603 clear_buddies(cfs_rq, se);
2605 if (se != cfs_rq->curr)
2606 __dequeue_entity(cfs_rq, se);
2608 account_entity_dequeue(cfs_rq, se);
2611 * Normalize the entity after updating the min_vruntime because the
2612 * update can refer to the ->curr item and we need to reflect this
2613 * movement in our normalized position.
2615 if (!(flags & DEQUEUE_SLEEP))
2616 se->vruntime -= cfs_rq->min_vruntime;
2618 /* return excess runtime on last dequeue */
2619 return_cfs_rq_runtime(cfs_rq);
2621 update_min_vruntime(cfs_rq);
2622 update_cfs_shares(cfs_rq);
2626 * Preempt the current task with a newly woken task if needed:
2629 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2631 unsigned long ideal_runtime, delta_exec;
2632 struct sched_entity *se;
2635 ideal_runtime = sched_slice(cfs_rq, curr);
2636 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2637 if (delta_exec > ideal_runtime) {
2638 resched_task(rq_of(cfs_rq)->curr);
2640 * The current task ran long enough, ensure it doesn't get
2641 * re-elected due to buddy favours.
2643 clear_buddies(cfs_rq, curr);
2648 * Ensure that a task that missed wakeup preemption by a
2649 * narrow margin doesn't have to wait for a full slice.
2650 * This also mitigates buddy induced latencies under load.
2652 if (delta_exec < sysctl_sched_min_granularity)
2655 se = __pick_first_entity(cfs_rq);
2656 delta = curr->vruntime - se->vruntime;
2661 if (delta > ideal_runtime)
2662 resched_task(rq_of(cfs_rq)->curr);
2666 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2668 /* 'current' is not kept within the tree. */
2671 * Any task has to be enqueued before it get to execute on
2672 * a CPU. So account for the time it spent waiting on the
2675 update_stats_wait_end(cfs_rq, se);
2676 __dequeue_entity(cfs_rq, se);
2679 update_stats_curr_start(cfs_rq, se);
2681 #ifdef CONFIG_SCHEDSTATS
2683 * Track our maximum slice length, if the CPU's load is at
2684 * least twice that of our own weight (i.e. dont track it
2685 * when there are only lesser-weight tasks around):
2687 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2688 se->statistics.slice_max = max(se->statistics.slice_max,
2689 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2692 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2696 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2699 * Pick the next process, keeping these things in mind, in this order:
2700 * 1) keep things fair between processes/task groups
2701 * 2) pick the "next" process, since someone really wants that to run
2702 * 3) pick the "last" process, for cache locality
2703 * 4) do not run the "skip" process, if something else is available
2705 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2707 struct sched_entity *se = __pick_first_entity(cfs_rq);
2708 struct sched_entity *left = se;
2711 * Avoid running the skip buddy, if running something else can
2712 * be done without getting too unfair.
2714 if (cfs_rq->skip == se) {
2715 struct sched_entity *second = __pick_next_entity(se);
2716 if (second && wakeup_preempt_entity(second, left) < 1)
2721 * Prefer last buddy, try to return the CPU to a preempted task.
2723 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2727 * Someone really wants this to run. If it's not unfair, run it.
2729 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2732 clear_buddies(cfs_rq, se);
2737 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2739 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2742 * If still on the runqueue then deactivate_task()
2743 * was not called and update_curr() has to be done:
2746 update_curr(cfs_rq);
2748 /* throttle cfs_rqs exceeding runtime */
2749 check_cfs_rq_runtime(cfs_rq);
2751 check_spread(cfs_rq, prev);
2753 update_stats_wait_start(cfs_rq, prev);
2754 /* Put 'current' back into the tree. */
2755 __enqueue_entity(cfs_rq, prev);
2756 /* in !on_rq case, update occurred at dequeue */
2757 update_entity_load_avg(prev, 1);
2759 cfs_rq->curr = NULL;
2763 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2766 * Update run-time statistics of the 'current'.
2768 update_curr(cfs_rq);
2771 * Ensure that runnable average is periodically updated.
2773 update_entity_load_avg(curr, 1);
2774 update_cfs_rq_blocked_load(cfs_rq, 1);
2775 update_cfs_shares(cfs_rq);
2777 #ifdef CONFIG_SCHED_HRTICK
2779 * queued ticks are scheduled to match the slice, so don't bother
2780 * validating it and just reschedule.
2783 resched_task(rq_of(cfs_rq)->curr);
2787 * don't let the period tick interfere with the hrtick preemption
2789 if (!sched_feat(DOUBLE_TICK) &&
2790 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2794 if (cfs_rq->nr_running > 1)
2795 check_preempt_tick(cfs_rq, curr);
2799 /**************************************************
2800 * CFS bandwidth control machinery
2803 #ifdef CONFIG_CFS_BANDWIDTH
2805 #ifdef HAVE_JUMP_LABEL
2806 static struct static_key __cfs_bandwidth_used;
2808 static inline bool cfs_bandwidth_used(void)
2810 return static_key_false(&__cfs_bandwidth_used);
2813 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2815 /* only need to count groups transitioning between enabled/!enabled */
2816 if (enabled && !was_enabled)
2817 static_key_slow_inc(&__cfs_bandwidth_used);
2818 else if (!enabled && was_enabled)
2819 static_key_slow_dec(&__cfs_bandwidth_used);
2821 #else /* HAVE_JUMP_LABEL */
2822 static bool cfs_bandwidth_used(void)
2827 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2828 #endif /* HAVE_JUMP_LABEL */
2831 * default period for cfs group bandwidth.
2832 * default: 0.1s, units: nanoseconds
2834 static inline u64 default_cfs_period(void)
2836 return 100000000ULL;
2839 static inline u64 sched_cfs_bandwidth_slice(void)
2841 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2845 * Replenish runtime according to assigned quota and update expiration time.
2846 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2847 * additional synchronization around rq->lock.
2849 * requires cfs_b->lock
2851 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2855 if (cfs_b->quota == RUNTIME_INF)
2858 now = sched_clock_cpu(smp_processor_id());
2859 cfs_b->runtime = cfs_b->quota;
2860 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2863 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2865 return &tg->cfs_bandwidth;
2868 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2869 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2871 if (unlikely(cfs_rq->throttle_count))
2872 return cfs_rq->throttled_clock_task;
2874 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2877 /* returns 0 on failure to allocate runtime */
2878 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2880 struct task_group *tg = cfs_rq->tg;
2881 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2882 u64 amount = 0, min_amount, expires;
2884 /* note: this is a positive sum as runtime_remaining <= 0 */
2885 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2887 raw_spin_lock(&cfs_b->lock);
2888 if (cfs_b->quota == RUNTIME_INF)
2889 amount = min_amount;
2892 * If the bandwidth pool has become inactive, then at least one
2893 * period must have elapsed since the last consumption.
2894 * Refresh the global state and ensure bandwidth timer becomes
2897 if (!cfs_b->timer_active) {
2898 __refill_cfs_bandwidth_runtime(cfs_b);
2899 __start_cfs_bandwidth(cfs_b);
2902 if (cfs_b->runtime > 0) {
2903 amount = min(cfs_b->runtime, min_amount);
2904 cfs_b->runtime -= amount;
2908 expires = cfs_b->runtime_expires;
2909 raw_spin_unlock(&cfs_b->lock);
2911 cfs_rq->runtime_remaining += amount;
2913 * we may have advanced our local expiration to account for allowed
2914 * spread between our sched_clock and the one on which runtime was
2917 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2918 cfs_rq->runtime_expires = expires;
2920 return cfs_rq->runtime_remaining > 0;
2924 * Note: This depends on the synchronization provided by sched_clock and the
2925 * fact that rq->clock snapshots this value.
2927 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2929 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2931 /* if the deadline is ahead of our clock, nothing to do */
2932 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2935 if (cfs_rq->runtime_remaining < 0)
2939 * If the local deadline has passed we have to consider the
2940 * possibility that our sched_clock is 'fast' and the global deadline
2941 * has not truly expired.
2943 * Fortunately we can check determine whether this the case by checking
2944 * whether the global deadline has advanced.
2947 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2948 /* extend local deadline, drift is bounded above by 2 ticks */
2949 cfs_rq->runtime_expires += TICK_NSEC;
2951 /* global deadline is ahead, expiration has passed */
2952 cfs_rq->runtime_remaining = 0;
2956 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2957 unsigned long delta_exec)
2959 /* dock delta_exec before expiring quota (as it could span periods) */
2960 cfs_rq->runtime_remaining -= delta_exec;
2961 expire_cfs_rq_runtime(cfs_rq);
2963 if (likely(cfs_rq->runtime_remaining > 0))
2967 * if we're unable to extend our runtime we resched so that the active
2968 * hierarchy can be throttled
2970 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2971 resched_task(rq_of(cfs_rq)->curr);
2974 static __always_inline
2975 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2977 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2980 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2983 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2985 return cfs_bandwidth_used() && cfs_rq->throttled;
2988 /* check whether cfs_rq, or any parent, is throttled */
2989 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2991 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2995 * Ensure that neither of the group entities corresponding to src_cpu or
2996 * dest_cpu are members of a throttled hierarchy when performing group
2997 * load-balance operations.
2999 static inline int throttled_lb_pair(struct task_group *tg,
3000 int src_cpu, int dest_cpu)
3002 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3004 src_cfs_rq = tg->cfs_rq[src_cpu];
3005 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3007 return throttled_hierarchy(src_cfs_rq) ||
3008 throttled_hierarchy(dest_cfs_rq);
3011 /* updated child weight may affect parent so we have to do this bottom up */
3012 static int tg_unthrottle_up(struct task_group *tg, void *data)
3014 struct rq *rq = data;
3015 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3017 cfs_rq->throttle_count--;
3019 if (!cfs_rq->throttle_count) {
3020 /* adjust cfs_rq_clock_task() */
3021 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3022 cfs_rq->throttled_clock_task;
3029 static int tg_throttle_down(struct task_group *tg, void *data)
3031 struct rq *rq = data;
3032 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3034 /* group is entering throttled state, stop time */
3035 if (!cfs_rq->throttle_count)
3036 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3037 cfs_rq->throttle_count++;
3042 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3044 struct rq *rq = rq_of(cfs_rq);
3045 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3046 struct sched_entity *se;
3047 long task_delta, dequeue = 1;
3049 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3051 /* freeze hierarchy runnable averages while throttled */
3053 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3056 task_delta = cfs_rq->h_nr_running;
3057 for_each_sched_entity(se) {
3058 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3059 /* throttled entity or throttle-on-deactivate */
3064 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3065 qcfs_rq->h_nr_running -= task_delta;
3067 if (qcfs_rq->load.weight)
3072 rq->nr_running -= task_delta;
3074 cfs_rq->throttled = 1;
3075 cfs_rq->throttled_clock = rq_clock(rq);
3076 raw_spin_lock(&cfs_b->lock);
3077 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3078 raw_spin_unlock(&cfs_b->lock);
3081 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3083 struct rq *rq = rq_of(cfs_rq);
3084 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3085 struct sched_entity *se;
3089 se = cfs_rq->tg->se[cpu_of(rq)];
3091 cfs_rq->throttled = 0;
3093 update_rq_clock(rq);
3095 raw_spin_lock(&cfs_b->lock);
3096 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3097 list_del_rcu(&cfs_rq->throttled_list);
3098 raw_spin_unlock(&cfs_b->lock);
3100 /* update hierarchical throttle state */
3101 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3103 if (!cfs_rq->load.weight)
3106 task_delta = cfs_rq->h_nr_running;
3107 for_each_sched_entity(se) {
3111 cfs_rq = cfs_rq_of(se);
3113 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3114 cfs_rq->h_nr_running += task_delta;
3116 if (cfs_rq_throttled(cfs_rq))
3121 rq->nr_running += task_delta;
3123 /* determine whether we need to wake up potentially idle cpu */
3124 if (rq->curr == rq->idle && rq->cfs.nr_running)
3125 resched_task(rq->curr);
3128 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3129 u64 remaining, u64 expires)
3131 struct cfs_rq *cfs_rq;
3132 u64 runtime = remaining;
3135 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3137 struct rq *rq = rq_of(cfs_rq);
3139 raw_spin_lock(&rq->lock);
3140 if (!cfs_rq_throttled(cfs_rq))
3143 runtime = -cfs_rq->runtime_remaining + 1;
3144 if (runtime > remaining)
3145 runtime = remaining;
3146 remaining -= runtime;
3148 cfs_rq->runtime_remaining += runtime;
3149 cfs_rq->runtime_expires = expires;
3151 /* we check whether we're throttled above */
3152 if (cfs_rq->runtime_remaining > 0)
3153 unthrottle_cfs_rq(cfs_rq);
3156 raw_spin_unlock(&rq->lock);
3167 * Responsible for refilling a task_group's bandwidth and unthrottling its
3168 * cfs_rqs as appropriate. If there has been no activity within the last
3169 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3170 * used to track this state.
3172 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3174 u64 runtime, runtime_expires;
3175 int idle = 1, throttled;
3177 raw_spin_lock(&cfs_b->lock);
3178 /* no need to continue the timer with no bandwidth constraint */
3179 if (cfs_b->quota == RUNTIME_INF)
3182 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3183 /* idle depends on !throttled (for the case of a large deficit) */
3184 idle = cfs_b->idle && !throttled;
3185 cfs_b->nr_periods += overrun;
3187 /* if we're going inactive then everything else can be deferred */
3191 __refill_cfs_bandwidth_runtime(cfs_b);
3194 /* mark as potentially idle for the upcoming period */
3199 /* account preceding periods in which throttling occurred */
3200 cfs_b->nr_throttled += overrun;
3203 * There are throttled entities so we must first use the new bandwidth
3204 * to unthrottle them before making it generally available. This
3205 * ensures that all existing debts will be paid before a new cfs_rq is
3208 runtime = cfs_b->runtime;
3209 runtime_expires = cfs_b->runtime_expires;
3213 * This check is repeated as we are holding onto the new bandwidth
3214 * while we unthrottle. This can potentially race with an unthrottled
3215 * group trying to acquire new bandwidth from the global pool.
3217 while (throttled && runtime > 0) {
3218 raw_spin_unlock(&cfs_b->lock);
3219 /* we can't nest cfs_b->lock while distributing bandwidth */
3220 runtime = distribute_cfs_runtime(cfs_b, runtime,
3222 raw_spin_lock(&cfs_b->lock);
3224 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3227 /* return (any) remaining runtime */
3228 cfs_b->runtime = runtime;
3230 * While we are ensured activity in the period following an
3231 * unthrottle, this also covers the case in which the new bandwidth is
3232 * insufficient to cover the existing bandwidth deficit. (Forcing the
3233 * timer to remain active while there are any throttled entities.)
3238 cfs_b->timer_active = 0;
3239 raw_spin_unlock(&cfs_b->lock);
3244 /* a cfs_rq won't donate quota below this amount */
3245 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3246 /* minimum remaining period time to redistribute slack quota */
3247 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3248 /* how long we wait to gather additional slack before distributing */
3249 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3251 /* are we near the end of the current quota period? */
3252 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3254 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3257 /* if the call-back is running a quota refresh is already occurring */
3258 if (hrtimer_callback_running(refresh_timer))
3261 /* is a quota refresh about to occur? */
3262 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3263 if (remaining < min_expire)
3269 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3271 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3273 /* if there's a quota refresh soon don't bother with slack */
3274 if (runtime_refresh_within(cfs_b, min_left))
3277 start_bandwidth_timer(&cfs_b->slack_timer,
3278 ns_to_ktime(cfs_bandwidth_slack_period));
3281 /* we know any runtime found here is valid as update_curr() precedes return */
3282 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3284 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3285 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3287 if (slack_runtime <= 0)
3290 raw_spin_lock(&cfs_b->lock);
3291 if (cfs_b->quota != RUNTIME_INF &&
3292 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3293 cfs_b->runtime += slack_runtime;
3295 /* we are under rq->lock, defer unthrottling using a timer */
3296 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3297 !list_empty(&cfs_b->throttled_cfs_rq))
3298 start_cfs_slack_bandwidth(cfs_b);
3300 raw_spin_unlock(&cfs_b->lock);
3302 /* even if it's not valid for return we don't want to try again */
3303 cfs_rq->runtime_remaining -= slack_runtime;
3306 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3308 if (!cfs_bandwidth_used())
3311 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3314 __return_cfs_rq_runtime(cfs_rq);
3318 * This is done with a timer (instead of inline with bandwidth return) since
3319 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3321 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3323 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3326 /* confirm we're still not at a refresh boundary */
3327 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3330 raw_spin_lock(&cfs_b->lock);
3331 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3332 runtime = cfs_b->runtime;
3335 expires = cfs_b->runtime_expires;
3336 raw_spin_unlock(&cfs_b->lock);
3341 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3343 raw_spin_lock(&cfs_b->lock);
3344 if (expires == cfs_b->runtime_expires)
3345 cfs_b->runtime = runtime;
3346 raw_spin_unlock(&cfs_b->lock);
3350 * When a group wakes up we want to make sure that its quota is not already
3351 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3352 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3354 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3356 if (!cfs_bandwidth_used())
3359 /* an active group must be handled by the update_curr()->put() path */
3360 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3363 /* ensure the group is not already throttled */
3364 if (cfs_rq_throttled(cfs_rq))
3367 /* update runtime allocation */
3368 account_cfs_rq_runtime(cfs_rq, 0);
3369 if (cfs_rq->runtime_remaining <= 0)
3370 throttle_cfs_rq(cfs_rq);
3373 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3374 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3376 if (!cfs_bandwidth_used())
3379 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3383 * it's possible for a throttled entity to be forced into a running
3384 * state (e.g. set_curr_task), in this case we're finished.
3386 if (cfs_rq_throttled(cfs_rq))
3389 throttle_cfs_rq(cfs_rq);
3392 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3394 struct cfs_bandwidth *cfs_b =
3395 container_of(timer, struct cfs_bandwidth, slack_timer);
3396 do_sched_cfs_slack_timer(cfs_b);
3398 return HRTIMER_NORESTART;
3401 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3403 struct cfs_bandwidth *cfs_b =
3404 container_of(timer, struct cfs_bandwidth, period_timer);
3410 now = hrtimer_cb_get_time(timer);
3411 overrun = hrtimer_forward(timer, now, cfs_b->period);
3416 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3419 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3422 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3424 raw_spin_lock_init(&cfs_b->lock);
3426 cfs_b->quota = RUNTIME_INF;
3427 cfs_b->period = ns_to_ktime(default_cfs_period());
3429 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3430 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3431 cfs_b->period_timer.function = sched_cfs_period_timer;
3432 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3433 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3436 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3438 cfs_rq->runtime_enabled = 0;
3439 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3442 /* requires cfs_b->lock, may release to reprogram timer */
3443 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3446 * The timer may be active because we're trying to set a new bandwidth
3447 * period or because we're racing with the tear-down path
3448 * (timer_active==0 becomes visible before the hrtimer call-back
3449 * terminates). In either case we ensure that it's re-programmed
3451 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3452 raw_spin_unlock(&cfs_b->lock);
3453 /* ensure cfs_b->lock is available while we wait */
3454 hrtimer_cancel(&cfs_b->period_timer);
3456 raw_spin_lock(&cfs_b->lock);
3457 /* if someone else restarted the timer then we're done */
3458 if (cfs_b->timer_active)
3462 cfs_b->timer_active = 1;
3463 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3466 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3468 hrtimer_cancel(&cfs_b->period_timer);
3469 hrtimer_cancel(&cfs_b->slack_timer);
3472 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3474 struct cfs_rq *cfs_rq;
3476 for_each_leaf_cfs_rq(rq, cfs_rq) {
3477 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3479 if (!cfs_rq->runtime_enabled)
3483 * clock_task is not advancing so we just need to make sure
3484 * there's some valid quota amount
3486 cfs_rq->runtime_remaining = cfs_b->quota;
3487 if (cfs_rq_throttled(cfs_rq))
3488 unthrottle_cfs_rq(cfs_rq);
3492 #else /* CONFIG_CFS_BANDWIDTH */
3493 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3495 return rq_clock_task(rq_of(cfs_rq));
3498 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3499 unsigned long delta_exec) {}
3500 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3501 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3502 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3504 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3509 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3514 static inline int throttled_lb_pair(struct task_group *tg,
3515 int src_cpu, int dest_cpu)
3520 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3522 #ifdef CONFIG_FAIR_GROUP_SCHED
3523 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3526 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3530 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3531 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3533 #endif /* CONFIG_CFS_BANDWIDTH */
3535 /**************************************************
3536 * CFS operations on tasks:
3539 #ifdef CONFIG_SCHED_HRTICK
3540 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3542 struct sched_entity *se = &p->se;
3543 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3545 WARN_ON(task_rq(p) != rq);
3547 if (cfs_rq->nr_running > 1) {
3548 u64 slice = sched_slice(cfs_rq, se);
3549 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3550 s64 delta = slice - ran;
3559 * Don't schedule slices shorter than 10000ns, that just
3560 * doesn't make sense. Rely on vruntime for fairness.
3563 delta = max_t(s64, 10000LL, delta);
3565 hrtick_start(rq, delta);
3570 * called from enqueue/dequeue and updates the hrtick when the
3571 * current task is from our class and nr_running is low enough
3574 static void hrtick_update(struct rq *rq)
3576 struct task_struct *curr = rq->curr;
3578 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3581 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3582 hrtick_start_fair(rq, curr);
3584 #else /* !CONFIG_SCHED_HRTICK */
3586 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3590 static inline void hrtick_update(struct rq *rq)
3596 * The enqueue_task method is called before nr_running is
3597 * increased. Here we update the fair scheduling stats and
3598 * then put the task into the rbtree:
3601 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3603 struct cfs_rq *cfs_rq;
3604 struct sched_entity *se = &p->se;
3606 for_each_sched_entity(se) {
3609 cfs_rq = cfs_rq_of(se);
3610 enqueue_entity(cfs_rq, se, flags);
3613 * end evaluation on encountering a throttled cfs_rq
3615 * note: in the case of encountering a throttled cfs_rq we will
3616 * post the final h_nr_running increment below.
3618 if (cfs_rq_throttled(cfs_rq))
3620 cfs_rq->h_nr_running++;
3622 flags = ENQUEUE_WAKEUP;
3625 for_each_sched_entity(se) {
3626 cfs_rq = cfs_rq_of(se);
3627 cfs_rq->h_nr_running++;
3629 if (cfs_rq_throttled(cfs_rq))
3632 update_cfs_shares(cfs_rq);
3633 update_entity_load_avg(se, 1);
3637 update_rq_runnable_avg(rq, rq->nr_running);
3643 static void set_next_buddy(struct sched_entity *se);
3646 * The dequeue_task method is called before nr_running is
3647 * decreased. We remove the task from the rbtree and
3648 * update the fair scheduling stats:
3650 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3652 struct cfs_rq *cfs_rq;
3653 struct sched_entity *se = &p->se;
3654 int task_sleep = flags & DEQUEUE_SLEEP;
3656 for_each_sched_entity(se) {
3657 cfs_rq = cfs_rq_of(se);
3658 dequeue_entity(cfs_rq, se, flags);
3661 * end evaluation on encountering a throttled cfs_rq
3663 * note: in the case of encountering a throttled cfs_rq we will
3664 * post the final h_nr_running decrement below.
3666 if (cfs_rq_throttled(cfs_rq))
3668 cfs_rq->h_nr_running--;
3670 /* Don't dequeue parent if it has other entities besides us */
3671 if (cfs_rq->load.weight) {
3673 * Bias pick_next to pick a task from this cfs_rq, as
3674 * p is sleeping when it is within its sched_slice.
3676 if (task_sleep && parent_entity(se))
3677 set_next_buddy(parent_entity(se));
3679 /* avoid re-evaluating load for this entity */
3680 se = parent_entity(se);
3683 flags |= DEQUEUE_SLEEP;
3686 for_each_sched_entity(se) {
3687 cfs_rq = cfs_rq_of(se);
3688 cfs_rq->h_nr_running--;
3690 if (cfs_rq_throttled(cfs_rq))
3693 update_cfs_shares(cfs_rq);
3694 update_entity_load_avg(se, 1);
3699 update_rq_runnable_avg(rq, 1);
3705 /* Used instead of source_load when we know the type == 0 */
3706 static unsigned long weighted_cpuload(const int cpu)
3708 return cpu_rq(cpu)->cfs.runnable_load_avg;
3712 * Return a low guess at the load of a migration-source cpu weighted
3713 * according to the scheduling class and "nice" value.
3715 * We want to under-estimate the load of migration sources, to
3716 * balance conservatively.
3718 static unsigned long source_load(int cpu, int type)
3720 struct rq *rq = cpu_rq(cpu);
3721 unsigned long total = weighted_cpuload(cpu);
3723 if (type == 0 || !sched_feat(LB_BIAS))
3726 return min(rq->cpu_load[type-1], total);
3730 * Return a high guess at the load of a migration-target cpu weighted
3731 * according to the scheduling class and "nice" value.
3733 static unsigned long target_load(int cpu, int type)
3735 struct rq *rq = cpu_rq(cpu);
3736 unsigned long total = weighted_cpuload(cpu);
3738 if (type == 0 || !sched_feat(LB_BIAS))
3741 return max(rq->cpu_load[type-1], total);
3744 static unsigned long power_of(int cpu)
3746 return cpu_rq(cpu)->cpu_power;
3749 static unsigned long cpu_avg_load_per_task(int cpu)
3751 struct rq *rq = cpu_rq(cpu);
3752 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3753 unsigned long load_avg = rq->cfs.runnable_load_avg;
3756 return load_avg / nr_running;
3761 static void record_wakee(struct task_struct *p)
3764 * Rough decay (wiping) for cost saving, don't worry
3765 * about the boundary, really active task won't care
3768 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3769 current->wakee_flips = 0;
3770 current->wakee_flip_decay_ts = jiffies;
3773 if (current->last_wakee != p) {
3774 current->last_wakee = p;
3775 current->wakee_flips++;
3779 static void task_waking_fair(struct task_struct *p)
3781 struct sched_entity *se = &p->se;
3782 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3785 #ifndef CONFIG_64BIT
3786 u64 min_vruntime_copy;
3789 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3791 min_vruntime = cfs_rq->min_vruntime;
3792 } while (min_vruntime != min_vruntime_copy);
3794 min_vruntime = cfs_rq->min_vruntime;
3797 se->vruntime -= min_vruntime;
3801 #ifdef CONFIG_FAIR_GROUP_SCHED
3803 * effective_load() calculates the load change as seen from the root_task_group
3805 * Adding load to a group doesn't make a group heavier, but can cause movement
3806 * of group shares between cpus. Assuming the shares were perfectly aligned one
3807 * can calculate the shift in shares.
3809 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3810 * on this @cpu and results in a total addition (subtraction) of @wg to the
3811 * total group weight.
3813 * Given a runqueue weight distribution (rw_i) we can compute a shares
3814 * distribution (s_i) using:
3816 * s_i = rw_i / \Sum rw_j (1)
3818 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3819 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3820 * shares distribution (s_i):
3822 * rw_i = { 2, 4, 1, 0 }
3823 * s_i = { 2/7, 4/7, 1/7, 0 }
3825 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3826 * task used to run on and the CPU the waker is running on), we need to
3827 * compute the effect of waking a task on either CPU and, in case of a sync
3828 * wakeup, compute the effect of the current task going to sleep.
3830 * So for a change of @wl to the local @cpu with an overall group weight change
3831 * of @wl we can compute the new shares distribution (s'_i) using:
3833 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3835 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3836 * differences in waking a task to CPU 0. The additional task changes the
3837 * weight and shares distributions like:
3839 * rw'_i = { 3, 4, 1, 0 }
3840 * s'_i = { 3/8, 4/8, 1/8, 0 }
3842 * We can then compute the difference in effective weight by using:
3844 * dw_i = S * (s'_i - s_i) (3)
3846 * Where 'S' is the group weight as seen by its parent.
3848 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3849 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3850 * 4/7) times the weight of the group.
3852 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3854 struct sched_entity *se = tg->se[cpu];
3856 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3859 for_each_sched_entity(se) {
3865 * W = @wg + \Sum rw_j
3867 W = wg + calc_tg_weight(tg, se->my_q);
3872 w = se->my_q->load.weight + wl;
3875 * wl = S * s'_i; see (2)
3878 wl = (w * tg->shares) / W;
3883 * Per the above, wl is the new se->load.weight value; since
3884 * those are clipped to [MIN_SHARES, ...) do so now. See
3885 * calc_cfs_shares().
3887 if (wl < MIN_SHARES)
3891 * wl = dw_i = S * (s'_i - s_i); see (3)
3893 wl -= se->load.weight;
3896 * Recursively apply this logic to all parent groups to compute
3897 * the final effective load change on the root group. Since
3898 * only the @tg group gets extra weight, all parent groups can
3899 * only redistribute existing shares. @wl is the shift in shares
3900 * resulting from this level per the above.
3909 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3916 static int wake_wide(struct task_struct *p)
3918 int factor = this_cpu_read(sd_llc_size);
3921 * Yeah, it's the switching-frequency, could means many wakee or
3922 * rapidly switch, use factor here will just help to automatically
3923 * adjust the loose-degree, so bigger node will lead to more pull.
3925 if (p->wakee_flips > factor) {
3927 * wakee is somewhat hot, it needs certain amount of cpu
3928 * resource, so if waker is far more hot, prefer to leave
3931 if (current->wakee_flips > (factor * p->wakee_flips))
3938 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3940 s64 this_load, load;
3941 int idx, this_cpu, prev_cpu;
3942 unsigned long tl_per_task;
3943 struct task_group *tg;
3944 unsigned long weight;
3948 * If we wake multiple tasks be careful to not bounce
3949 * ourselves around too much.
3955 this_cpu = smp_processor_id();
3956 prev_cpu = task_cpu(p);
3957 load = source_load(prev_cpu, idx);
3958 this_load = target_load(this_cpu, idx);
3961 * If sync wakeup then subtract the (maximum possible)
3962 * effect of the currently running task from the load
3963 * of the current CPU:
3966 tg = task_group(current);
3967 weight = current->se.load.weight;
3969 this_load += effective_load(tg, this_cpu, -weight, -weight);
3970 load += effective_load(tg, prev_cpu, 0, -weight);
3974 weight = p->se.load.weight;
3977 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3978 * due to the sync cause above having dropped this_load to 0, we'll
3979 * always have an imbalance, but there's really nothing you can do
3980 * about that, so that's good too.
3982 * Otherwise check if either cpus are near enough in load to allow this
3983 * task to be woken on this_cpu.
3985 if (this_load > 0) {
3986 s64 this_eff_load, prev_eff_load;
3988 this_eff_load = 100;
3989 this_eff_load *= power_of(prev_cpu);
3990 this_eff_load *= this_load +
3991 effective_load(tg, this_cpu, weight, weight);
3993 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3994 prev_eff_load *= power_of(this_cpu);
3995 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3997 balanced = this_eff_load <= prev_eff_load;
4002 * If the currently running task will sleep within
4003 * a reasonable amount of time then attract this newly
4006 if (sync && balanced)
4009 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4010 tl_per_task = cpu_avg_load_per_task(this_cpu);
4013 (this_load <= load &&
4014 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4016 * This domain has SD_WAKE_AFFINE and
4017 * p is cache cold in this domain, and
4018 * there is no bad imbalance.
4020 schedstat_inc(sd, ttwu_move_affine);
4021 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4029 * find_idlest_group finds and returns the least busy CPU group within the
4032 static struct sched_group *
4033 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4034 int this_cpu, int load_idx)
4036 struct sched_group *idlest = NULL, *group = sd->groups;
4037 unsigned long min_load = ULONG_MAX, this_load = 0;
4038 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4041 unsigned long load, avg_load;
4045 /* Skip over this group if it has no CPUs allowed */
4046 if (!cpumask_intersects(sched_group_cpus(group),
4047 tsk_cpus_allowed(p)))
4050 local_group = cpumask_test_cpu(this_cpu,
4051 sched_group_cpus(group));
4053 /* Tally up the load of all CPUs in the group */
4056 for_each_cpu(i, sched_group_cpus(group)) {
4057 /* Bias balancing toward cpus of our domain */
4059 load = source_load(i, load_idx);
4061 load = target_load(i, load_idx);
4066 /* Adjust by relative CPU power of the group */
4067 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4070 this_load = avg_load;
4071 } else if (avg_load < min_load) {
4072 min_load = avg_load;
4075 } while (group = group->next, group != sd->groups);
4077 if (!idlest || 100*this_load < imbalance*min_load)
4083 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4086 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4088 unsigned long load, min_load = ULONG_MAX;
4092 /* Traverse only the allowed CPUs */
4093 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4094 load = weighted_cpuload(i);
4096 if (load < min_load || (load == min_load && i == this_cpu)) {
4106 * Try and locate an idle CPU in the sched_domain.
4108 static int select_idle_sibling(struct task_struct *p, int target)
4110 struct sched_domain *sd;
4111 struct sched_group *sg;
4112 int i = task_cpu(p);
4114 if (idle_cpu(target))
4118 * If the prevous cpu is cache affine and idle, don't be stupid.
4120 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4124 * Otherwise, iterate the domains and find an elegible idle cpu.
4126 sd = rcu_dereference(per_cpu(sd_llc, target));
4127 for_each_lower_domain(sd) {
4130 if (!cpumask_intersects(sched_group_cpus(sg),
4131 tsk_cpus_allowed(p)))
4134 for_each_cpu(i, sched_group_cpus(sg)) {
4135 if (i == target || !idle_cpu(i))
4139 target = cpumask_first_and(sched_group_cpus(sg),
4140 tsk_cpus_allowed(p));
4144 } while (sg != sd->groups);
4151 * sched_balance_self: balance the current task (running on cpu) in domains
4152 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4155 * Balance, ie. select the least loaded group.
4157 * Returns the target CPU number, or the same CPU if no balancing is needed.
4159 * preempt must be disabled.
4162 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4164 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4165 int cpu = smp_processor_id();
4167 int want_affine = 0;
4168 int sync = wake_flags & WF_SYNC;
4170 if (p->nr_cpus_allowed == 1)
4173 if (sd_flag & SD_BALANCE_WAKE) {
4174 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4180 for_each_domain(cpu, tmp) {
4181 if (!(tmp->flags & SD_LOAD_BALANCE))
4185 * If both cpu and prev_cpu are part of this domain,
4186 * cpu is a valid SD_WAKE_AFFINE target.
4188 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4189 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4194 if (tmp->flags & sd_flag)
4199 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4202 new_cpu = select_idle_sibling(p, prev_cpu);
4207 int load_idx = sd->forkexec_idx;
4208 struct sched_group *group;
4211 if (!(sd->flags & sd_flag)) {
4216 if (sd_flag & SD_BALANCE_WAKE)
4217 load_idx = sd->wake_idx;
4219 group = find_idlest_group(sd, p, cpu, load_idx);
4225 new_cpu = find_idlest_cpu(group, p, cpu);
4226 if (new_cpu == -1 || new_cpu == cpu) {
4227 /* Now try balancing at a lower domain level of cpu */
4232 /* Now try balancing at a lower domain level of new_cpu */
4234 weight = sd->span_weight;
4236 for_each_domain(cpu, tmp) {
4237 if (weight <= tmp->span_weight)
4239 if (tmp->flags & sd_flag)
4242 /* while loop will break here if sd == NULL */
4251 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4252 * cfs_rq_of(p) references at time of call are still valid and identify the
4253 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4254 * other assumptions, including the state of rq->lock, should be made.
4257 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4259 struct sched_entity *se = &p->se;
4260 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4263 * Load tracking: accumulate removed load so that it can be processed
4264 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4265 * to blocked load iff they have a positive decay-count. It can never
4266 * be negative here since on-rq tasks have decay-count == 0.
4268 if (se->avg.decay_count) {
4269 se->avg.decay_count = -__synchronize_entity_decay(se);
4270 atomic_long_add(se->avg.load_avg_contrib,
4271 &cfs_rq->removed_load);
4274 #endif /* CONFIG_SMP */
4276 static unsigned long
4277 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4279 unsigned long gran = sysctl_sched_wakeup_granularity;
4282 * Since its curr running now, convert the gran from real-time
4283 * to virtual-time in his units.
4285 * By using 'se' instead of 'curr' we penalize light tasks, so
4286 * they get preempted easier. That is, if 'se' < 'curr' then
4287 * the resulting gran will be larger, therefore penalizing the
4288 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4289 * be smaller, again penalizing the lighter task.
4291 * This is especially important for buddies when the leftmost
4292 * task is higher priority than the buddy.
4294 return calc_delta_fair(gran, se);
4298 * Should 'se' preempt 'curr'.
4312 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4314 s64 gran, vdiff = curr->vruntime - se->vruntime;
4319 gran = wakeup_gran(curr, se);
4326 static void set_last_buddy(struct sched_entity *se)
4328 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4331 for_each_sched_entity(se)
4332 cfs_rq_of(se)->last = se;
4335 static void set_next_buddy(struct sched_entity *se)
4337 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4340 for_each_sched_entity(se)
4341 cfs_rq_of(se)->next = se;
4344 static void set_skip_buddy(struct sched_entity *se)
4346 for_each_sched_entity(se)
4347 cfs_rq_of(se)->skip = se;
4351 * Preempt the current task with a newly woken task if needed:
4353 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4355 struct task_struct *curr = rq->curr;
4356 struct sched_entity *se = &curr->se, *pse = &p->se;
4357 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4358 int scale = cfs_rq->nr_running >= sched_nr_latency;
4359 int next_buddy_marked = 0;
4361 if (unlikely(se == pse))
4365 * This is possible from callers such as move_task(), in which we
4366 * unconditionally check_prempt_curr() after an enqueue (which may have
4367 * lead to a throttle). This both saves work and prevents false
4368 * next-buddy nomination below.
4370 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4373 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4374 set_next_buddy(pse);
4375 next_buddy_marked = 1;
4379 * We can come here with TIF_NEED_RESCHED already set from new task
4382 * Note: this also catches the edge-case of curr being in a throttled
4383 * group (e.g. via set_curr_task), since update_curr() (in the
4384 * enqueue of curr) will have resulted in resched being set. This
4385 * prevents us from potentially nominating it as a false LAST_BUDDY
4388 if (test_tsk_need_resched(curr))
4391 /* Idle tasks are by definition preempted by non-idle tasks. */
4392 if (unlikely(curr->policy == SCHED_IDLE) &&
4393 likely(p->policy != SCHED_IDLE))
4397 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4398 * is driven by the tick):
4400 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4403 find_matching_se(&se, &pse);
4404 update_curr(cfs_rq_of(se));
4406 if (wakeup_preempt_entity(se, pse) == 1) {
4408 * Bias pick_next to pick the sched entity that is
4409 * triggering this preemption.
4411 if (!next_buddy_marked)
4412 set_next_buddy(pse);
4421 * Only set the backward buddy when the current task is still
4422 * on the rq. This can happen when a wakeup gets interleaved
4423 * with schedule on the ->pre_schedule() or idle_balance()
4424 * point, either of which can * drop the rq lock.
4426 * Also, during early boot the idle thread is in the fair class,
4427 * for obvious reasons its a bad idea to schedule back to it.
4429 if (unlikely(!se->on_rq || curr == rq->idle))
4432 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4436 static struct task_struct *pick_next_task_fair(struct rq *rq)
4438 struct task_struct *p;
4439 struct cfs_rq *cfs_rq = &rq->cfs;
4440 struct sched_entity *se;
4442 if (!cfs_rq->nr_running)
4446 se = pick_next_entity(cfs_rq);
4447 set_next_entity(cfs_rq, se);
4448 cfs_rq = group_cfs_rq(se);
4452 if (hrtick_enabled(rq))
4453 hrtick_start_fair(rq, p);
4459 * Account for a descheduled task:
4461 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4463 struct sched_entity *se = &prev->se;
4464 struct cfs_rq *cfs_rq;
4466 for_each_sched_entity(se) {
4467 cfs_rq = cfs_rq_of(se);
4468 put_prev_entity(cfs_rq, se);
4473 * sched_yield() is very simple
4475 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4477 static void yield_task_fair(struct rq *rq)
4479 struct task_struct *curr = rq->curr;
4480 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4481 struct sched_entity *se = &curr->se;
4484 * Are we the only task in the tree?
4486 if (unlikely(rq->nr_running == 1))
4489 clear_buddies(cfs_rq, se);
4491 if (curr->policy != SCHED_BATCH) {
4492 update_rq_clock(rq);
4494 * Update run-time statistics of the 'current'.
4496 update_curr(cfs_rq);
4498 * Tell update_rq_clock() that we've just updated,
4499 * so we don't do microscopic update in schedule()
4500 * and double the fastpath cost.
4502 rq->skip_clock_update = 1;
4508 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4510 struct sched_entity *se = &p->se;
4512 /* throttled hierarchies are not runnable */
4513 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4516 /* Tell the scheduler that we'd really like pse to run next. */
4519 yield_task_fair(rq);
4525 /**************************************************
4526 * Fair scheduling class load-balancing methods.
4530 * The purpose of load-balancing is to achieve the same basic fairness the
4531 * per-cpu scheduler provides, namely provide a proportional amount of compute
4532 * time to each task. This is expressed in the following equation:
4534 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4536 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4537 * W_i,0 is defined as:
4539 * W_i,0 = \Sum_j w_i,j (2)
4541 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4542 * is derived from the nice value as per prio_to_weight[].
4544 * The weight average is an exponential decay average of the instantaneous
4547 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4549 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4550 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4551 * can also include other factors [XXX].
4553 * To achieve this balance we define a measure of imbalance which follows
4554 * directly from (1):
4556 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4558 * We them move tasks around to minimize the imbalance. In the continuous
4559 * function space it is obvious this converges, in the discrete case we get
4560 * a few fun cases generally called infeasible weight scenarios.
4563 * - infeasible weights;
4564 * - local vs global optima in the discrete case. ]
4569 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4570 * for all i,j solution, we create a tree of cpus that follows the hardware
4571 * topology where each level pairs two lower groups (or better). This results
4572 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4573 * tree to only the first of the previous level and we decrease the frequency
4574 * of load-balance at each level inv. proportional to the number of cpus in
4580 * \Sum { --- * --- * 2^i } = O(n) (5)
4582 * `- size of each group
4583 * | | `- number of cpus doing load-balance
4585 * `- sum over all levels
4587 * Coupled with a limit on how many tasks we can migrate every balance pass,
4588 * this makes (5) the runtime complexity of the balancer.
4590 * An important property here is that each CPU is still (indirectly) connected
4591 * to every other cpu in at most O(log n) steps:
4593 * The adjacency matrix of the resulting graph is given by:
4596 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4599 * And you'll find that:
4601 * A^(log_2 n)_i,j != 0 for all i,j (7)
4603 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4604 * The task movement gives a factor of O(m), giving a convergence complexity
4607 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4612 * In order to avoid CPUs going idle while there's still work to do, new idle
4613 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4614 * tree itself instead of relying on other CPUs to bring it work.
4616 * This adds some complexity to both (5) and (8) but it reduces the total idle
4624 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4627 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4632 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4634 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4636 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4639 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4640 * rewrite all of this once again.]
4643 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4645 enum fbq_type { regular, remote, all };
4647 #define LBF_ALL_PINNED 0x01
4648 #define LBF_NEED_BREAK 0x02
4649 #define LBF_DST_PINNED 0x04
4650 #define LBF_SOME_PINNED 0x08
4653 struct sched_domain *sd;
4661 struct cpumask *dst_grpmask;
4663 enum cpu_idle_type idle;
4665 /* The set of CPUs under consideration for load-balancing */
4666 struct cpumask *cpus;
4671 unsigned int loop_break;
4672 unsigned int loop_max;
4674 enum fbq_type fbq_type;
4678 * move_task - move a task from one runqueue to another runqueue.
4679 * Both runqueues must be locked.
4681 static void move_task(struct task_struct *p, struct lb_env *env)
4683 deactivate_task(env->src_rq, p, 0);
4684 set_task_cpu(p, env->dst_cpu);
4685 activate_task(env->dst_rq, p, 0);
4686 check_preempt_curr(env->dst_rq, p, 0);
4687 #ifdef CONFIG_NUMA_BALANCING
4688 if (p->numa_preferred_nid != -1) {
4689 int src_nid = cpu_to_node(env->src_cpu);
4690 int dst_nid = cpu_to_node(env->dst_cpu);
4693 * If the load balancer has moved the task then limit
4694 * migrations from taking place in the short term in
4695 * case this is a short-lived migration.
4697 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4698 p->numa_migrate_seq = 0;
4704 * Is this task likely cache-hot:
4707 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4711 if (p->sched_class != &fair_sched_class)
4714 if (unlikely(p->policy == SCHED_IDLE))
4718 * Buddy candidates are cache hot:
4720 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4721 (&p->se == cfs_rq_of(&p->se)->next ||
4722 &p->se == cfs_rq_of(&p->se)->last))
4725 if (sysctl_sched_migration_cost == -1)
4727 if (sysctl_sched_migration_cost == 0)
4730 delta = now - p->se.exec_start;
4732 return delta < (s64)sysctl_sched_migration_cost;
4735 #ifdef CONFIG_NUMA_BALANCING
4736 /* Returns true if the destination node has incurred more faults */
4737 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4739 int src_nid, dst_nid;
4741 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4742 !(env->sd->flags & SD_NUMA)) {
4746 src_nid = cpu_to_node(env->src_cpu);
4747 dst_nid = cpu_to_node(env->dst_cpu);
4749 if (src_nid == dst_nid)
4752 /* Always encourage migration to the preferred node. */
4753 if (dst_nid == p->numa_preferred_nid)
4756 /* If both task and group weight improve, this move is a winner. */
4757 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4758 group_weight(p, dst_nid) > group_weight(p, src_nid))
4765 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4767 int src_nid, dst_nid;
4769 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4772 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4775 src_nid = cpu_to_node(env->src_cpu);
4776 dst_nid = cpu_to_node(env->dst_cpu);
4778 if (src_nid == dst_nid)
4781 /* Migrating away from the preferred node is always bad. */
4782 if (src_nid == p->numa_preferred_nid)
4785 /* If either task or group weight get worse, don't do it. */
4786 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4787 group_weight(p, dst_nid) < group_weight(p, src_nid))
4794 static inline bool migrate_improves_locality(struct task_struct *p,
4800 static inline bool migrate_degrades_locality(struct task_struct *p,
4808 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4811 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4813 int tsk_cache_hot = 0;
4815 * We do not migrate tasks that are:
4816 * 1) throttled_lb_pair, or
4817 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4818 * 3) running (obviously), or
4819 * 4) are cache-hot on their current CPU.
4821 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4824 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4827 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4829 env->flags |= LBF_SOME_PINNED;
4832 * Remember if this task can be migrated to any other cpu in
4833 * our sched_group. We may want to revisit it if we couldn't
4834 * meet load balance goals by pulling other tasks on src_cpu.
4836 * Also avoid computing new_dst_cpu if we have already computed
4837 * one in current iteration.
4839 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4842 /* Prevent to re-select dst_cpu via env's cpus */
4843 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4844 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4845 env->flags |= LBF_DST_PINNED;
4846 env->new_dst_cpu = cpu;
4854 /* Record that we found atleast one task that could run on dst_cpu */
4855 env->flags &= ~LBF_ALL_PINNED;
4857 if (task_running(env->src_rq, p)) {
4858 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4863 * Aggressive migration if:
4864 * 1) destination numa is preferred
4865 * 2) task is cache cold, or
4866 * 3) too many balance attempts have failed.
4868 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4870 tsk_cache_hot = migrate_degrades_locality(p, env);
4872 if (migrate_improves_locality(p, env)) {
4873 #ifdef CONFIG_SCHEDSTATS
4874 if (tsk_cache_hot) {
4875 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4876 schedstat_inc(p, se.statistics.nr_forced_migrations);
4882 if (!tsk_cache_hot ||
4883 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4885 if (tsk_cache_hot) {
4886 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4887 schedstat_inc(p, se.statistics.nr_forced_migrations);
4893 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4898 * move_one_task tries to move exactly one task from busiest to this_rq, as
4899 * part of active balancing operations within "domain".
4900 * Returns 1 if successful and 0 otherwise.
4902 * Called with both runqueues locked.
4904 static int move_one_task(struct lb_env *env)
4906 struct task_struct *p, *n;
4908 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4909 if (!can_migrate_task(p, env))
4914 * Right now, this is only the second place move_task()
4915 * is called, so we can safely collect move_task()
4916 * stats here rather than inside move_task().
4918 schedstat_inc(env->sd, lb_gained[env->idle]);
4924 static const unsigned int sched_nr_migrate_break = 32;
4927 * move_tasks tries to move up to imbalance weighted load from busiest to
4928 * this_rq, as part of a balancing operation within domain "sd".
4929 * Returns 1 if successful and 0 otherwise.
4931 * Called with both runqueues locked.
4933 static int move_tasks(struct lb_env *env)
4935 struct list_head *tasks = &env->src_rq->cfs_tasks;
4936 struct task_struct *p;
4940 if (env->imbalance <= 0)
4943 while (!list_empty(tasks)) {
4944 p = list_first_entry(tasks, struct task_struct, se.group_node);
4947 /* We've more or less seen every task there is, call it quits */
4948 if (env->loop > env->loop_max)
4951 /* take a breather every nr_migrate tasks */
4952 if (env->loop > env->loop_break) {
4953 env->loop_break += sched_nr_migrate_break;
4954 env->flags |= LBF_NEED_BREAK;
4958 if (!can_migrate_task(p, env))
4961 load = task_h_load(p);
4963 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4966 if ((load / 2) > env->imbalance)
4971 env->imbalance -= load;
4973 #ifdef CONFIG_PREEMPT
4975 * NEWIDLE balancing is a source of latency, so preemptible
4976 * kernels will stop after the first task is pulled to minimize
4977 * the critical section.
4979 if (env->idle == CPU_NEWLY_IDLE)
4984 * We only want to steal up to the prescribed amount of
4987 if (env->imbalance <= 0)
4992 list_move_tail(&p->se.group_node, tasks);
4996 * Right now, this is one of only two places move_task() is called,
4997 * so we can safely collect move_task() stats here rather than
4998 * inside move_task().
5000 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5005 #ifdef CONFIG_FAIR_GROUP_SCHED
5007 * update tg->load_weight by folding this cpu's load_avg
5009 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5011 struct sched_entity *se = tg->se[cpu];
5012 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5014 /* throttled entities do not contribute to load */
5015 if (throttled_hierarchy(cfs_rq))
5018 update_cfs_rq_blocked_load(cfs_rq, 1);
5021 update_entity_load_avg(se, 1);
5023 * We pivot on our runnable average having decayed to zero for
5024 * list removal. This generally implies that all our children
5025 * have also been removed (modulo rounding error or bandwidth
5026 * control); however, such cases are rare and we can fix these
5029 * TODO: fix up out-of-order children on enqueue.
5031 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5032 list_del_leaf_cfs_rq(cfs_rq);
5034 struct rq *rq = rq_of(cfs_rq);
5035 update_rq_runnable_avg(rq, rq->nr_running);
5039 static void update_blocked_averages(int cpu)
5041 struct rq *rq = cpu_rq(cpu);
5042 struct cfs_rq *cfs_rq;
5043 unsigned long flags;
5045 raw_spin_lock_irqsave(&rq->lock, flags);
5046 update_rq_clock(rq);
5048 * Iterates the task_group tree in a bottom up fashion, see
5049 * list_add_leaf_cfs_rq() for details.
5051 for_each_leaf_cfs_rq(rq, cfs_rq) {
5053 * Note: We may want to consider periodically releasing
5054 * rq->lock about these updates so that creating many task
5055 * groups does not result in continually extending hold time.
5057 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5060 raw_spin_unlock_irqrestore(&rq->lock, flags);
5064 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5065 * This needs to be done in a top-down fashion because the load of a child
5066 * group is a fraction of its parents load.
5068 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5070 struct rq *rq = rq_of(cfs_rq);
5071 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5072 unsigned long now = jiffies;
5075 if (cfs_rq->last_h_load_update == now)
5078 cfs_rq->h_load_next = NULL;
5079 for_each_sched_entity(se) {
5080 cfs_rq = cfs_rq_of(se);
5081 cfs_rq->h_load_next = se;
5082 if (cfs_rq->last_h_load_update == now)
5087 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5088 cfs_rq->last_h_load_update = now;
5091 while ((se = cfs_rq->h_load_next) != NULL) {
5092 load = cfs_rq->h_load;
5093 load = div64_ul(load * se->avg.load_avg_contrib,
5094 cfs_rq->runnable_load_avg + 1);
5095 cfs_rq = group_cfs_rq(se);
5096 cfs_rq->h_load = load;
5097 cfs_rq->last_h_load_update = now;
5101 static unsigned long task_h_load(struct task_struct *p)
5103 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5105 update_cfs_rq_h_load(cfs_rq);
5106 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5107 cfs_rq->runnable_load_avg + 1);
5110 static inline void update_blocked_averages(int cpu)
5114 static unsigned long task_h_load(struct task_struct *p)
5116 return p->se.avg.load_avg_contrib;
5120 /********** Helpers for find_busiest_group ************************/
5122 * sg_lb_stats - stats of a sched_group required for load_balancing
5124 struct sg_lb_stats {
5125 unsigned long avg_load; /*Avg load across the CPUs of the group */
5126 unsigned long group_load; /* Total load over the CPUs of the group */
5127 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5128 unsigned long load_per_task;
5129 unsigned long group_power;
5130 unsigned int sum_nr_running; /* Nr tasks running in the group */
5131 unsigned int group_capacity;
5132 unsigned int idle_cpus;
5133 unsigned int group_weight;
5134 int group_imb; /* Is there an imbalance in the group ? */
5135 int group_has_capacity; /* Is there extra capacity in the group? */
5136 #ifdef CONFIG_NUMA_BALANCING
5137 unsigned int nr_numa_running;
5138 unsigned int nr_preferred_running;
5143 * sd_lb_stats - Structure to store the statistics of a sched_domain
5144 * during load balancing.
5146 struct sd_lb_stats {
5147 struct sched_group *busiest; /* Busiest group in this sd */
5148 struct sched_group *local; /* Local group in this sd */
5149 unsigned long total_load; /* Total load of all groups in sd */
5150 unsigned long total_pwr; /* Total power of all groups in sd */
5151 unsigned long avg_load; /* Average load across all groups in sd */
5153 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5154 struct sg_lb_stats local_stat; /* Statistics of the local group */
5157 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5160 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5161 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5162 * We must however clear busiest_stat::avg_load because
5163 * update_sd_pick_busiest() reads this before assignment.
5165 *sds = (struct sd_lb_stats){
5177 * get_sd_load_idx - Obtain the load index for a given sched domain.
5178 * @sd: The sched_domain whose load_idx is to be obtained.
5179 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5181 * Return: The load index.
5183 static inline int get_sd_load_idx(struct sched_domain *sd,
5184 enum cpu_idle_type idle)
5190 load_idx = sd->busy_idx;
5193 case CPU_NEWLY_IDLE:
5194 load_idx = sd->newidle_idx;
5197 load_idx = sd->idle_idx;
5204 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5206 return SCHED_POWER_SCALE;
5209 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5211 return default_scale_freq_power(sd, cpu);
5214 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5216 unsigned long weight = sd->span_weight;
5217 unsigned long smt_gain = sd->smt_gain;
5224 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5226 return default_scale_smt_power(sd, cpu);
5229 static unsigned long scale_rt_power(int cpu)
5231 struct rq *rq = cpu_rq(cpu);
5232 u64 total, available, age_stamp, avg;
5235 * Since we're reading these variables without serialization make sure
5236 * we read them once before doing sanity checks on them.
5238 age_stamp = ACCESS_ONCE(rq->age_stamp);
5239 avg = ACCESS_ONCE(rq->rt_avg);
5241 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5243 if (unlikely(total < avg)) {
5244 /* Ensures that power won't end up being negative */
5247 available = total - avg;
5250 if (unlikely((s64)total < SCHED_POWER_SCALE))
5251 total = SCHED_POWER_SCALE;
5253 total >>= SCHED_POWER_SHIFT;
5255 return div_u64(available, total);
5258 static void update_cpu_power(struct sched_domain *sd, int cpu)
5260 unsigned long weight = sd->span_weight;
5261 unsigned long power = SCHED_POWER_SCALE;
5262 struct sched_group *sdg = sd->groups;
5264 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5265 if (sched_feat(ARCH_POWER))
5266 power *= arch_scale_smt_power(sd, cpu);
5268 power *= default_scale_smt_power(sd, cpu);
5270 power >>= SCHED_POWER_SHIFT;
5273 sdg->sgp->power_orig = power;
5275 if (sched_feat(ARCH_POWER))
5276 power *= arch_scale_freq_power(sd, cpu);
5278 power *= default_scale_freq_power(sd, cpu);
5280 power >>= SCHED_POWER_SHIFT;
5282 power *= scale_rt_power(cpu);
5283 power >>= SCHED_POWER_SHIFT;
5288 cpu_rq(cpu)->cpu_power = power;
5289 sdg->sgp->power = power;
5292 void update_group_power(struct sched_domain *sd, int cpu)
5294 struct sched_domain *child = sd->child;
5295 struct sched_group *group, *sdg = sd->groups;
5296 unsigned long power, power_orig;
5297 unsigned long interval;
5299 interval = msecs_to_jiffies(sd->balance_interval);
5300 interval = clamp(interval, 1UL, max_load_balance_interval);
5301 sdg->sgp->next_update = jiffies + interval;
5304 update_cpu_power(sd, cpu);
5308 power_orig = power = 0;
5310 if (child->flags & SD_OVERLAP) {
5312 * SD_OVERLAP domains cannot assume that child groups
5313 * span the current group.
5316 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5317 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5319 power_orig += sg->sgp->power_orig;
5320 power += sg->sgp->power;
5324 * !SD_OVERLAP domains can assume that child groups
5325 * span the current group.
5328 group = child->groups;
5330 power_orig += group->sgp->power_orig;
5331 power += group->sgp->power;
5332 group = group->next;
5333 } while (group != child->groups);
5336 sdg->sgp->power_orig = power_orig;
5337 sdg->sgp->power = power;
5341 * Try and fix up capacity for tiny siblings, this is needed when
5342 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5343 * which on its own isn't powerful enough.
5345 * See update_sd_pick_busiest() and check_asym_packing().
5348 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5351 * Only siblings can have significantly less than SCHED_POWER_SCALE
5353 if (!(sd->flags & SD_SHARE_CPUPOWER))
5357 * If ~90% of the cpu_power is still there, we're good.
5359 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5366 * Group imbalance indicates (and tries to solve) the problem where balancing
5367 * groups is inadequate due to tsk_cpus_allowed() constraints.
5369 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5370 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5373 * { 0 1 2 3 } { 4 5 6 7 }
5376 * If we were to balance group-wise we'd place two tasks in the first group and
5377 * two tasks in the second group. Clearly this is undesired as it will overload
5378 * cpu 3 and leave one of the cpus in the second group unused.
5380 * The current solution to this issue is detecting the skew in the first group
5381 * by noticing the lower domain failed to reach balance and had difficulty
5382 * moving tasks due to affinity constraints.
5384 * When this is so detected; this group becomes a candidate for busiest; see
5385 * update_sd_pick_busiest(). And calculcate_imbalance() and
5386 * find_busiest_group() avoid some of the usual balance conditions to allow it
5387 * to create an effective group imbalance.
5389 * This is a somewhat tricky proposition since the next run might not find the
5390 * group imbalance and decide the groups need to be balanced again. A most
5391 * subtle and fragile situation.
5394 static inline int sg_imbalanced(struct sched_group *group)
5396 return group->sgp->imbalance;
5400 * Compute the group capacity.
5402 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5403 * first dividing out the smt factor and computing the actual number of cores
5404 * and limit power unit capacity with that.
5406 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5408 unsigned int capacity, smt, cpus;
5409 unsigned int power, power_orig;
5411 power = group->sgp->power;
5412 power_orig = group->sgp->power_orig;
5413 cpus = group->group_weight;
5415 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5416 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5417 capacity = cpus / smt; /* cores */
5419 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5421 capacity = fix_small_capacity(env->sd, group);
5427 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5428 * @env: The load balancing environment.
5429 * @group: sched_group whose statistics are to be updated.
5430 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5431 * @local_group: Does group contain this_cpu.
5432 * @sgs: variable to hold the statistics for this group.
5434 static inline void update_sg_lb_stats(struct lb_env *env,
5435 struct sched_group *group, int load_idx,
5436 int local_group, struct sg_lb_stats *sgs)
5438 unsigned long nr_running;
5442 memset(sgs, 0, sizeof(*sgs));
5444 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5445 struct rq *rq = cpu_rq(i);
5447 nr_running = rq->nr_running;
5449 /* Bias balancing toward cpus of our domain */
5451 load = target_load(i, load_idx);
5453 load = source_load(i, load_idx);
5455 sgs->group_load += load;
5456 sgs->sum_nr_running += nr_running;
5457 #ifdef CONFIG_NUMA_BALANCING
5458 sgs->nr_numa_running += rq->nr_numa_running;
5459 sgs->nr_preferred_running += rq->nr_preferred_running;
5461 sgs->sum_weighted_load += weighted_cpuload(i);
5466 /* Adjust by relative CPU power of the group */
5467 sgs->group_power = group->sgp->power;
5468 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5470 if (sgs->sum_nr_running)
5471 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5473 sgs->group_weight = group->group_weight;
5475 sgs->group_imb = sg_imbalanced(group);
5476 sgs->group_capacity = sg_capacity(env, group);
5478 if (sgs->group_capacity > sgs->sum_nr_running)
5479 sgs->group_has_capacity = 1;
5483 * update_sd_pick_busiest - return 1 on busiest group
5484 * @env: The load balancing environment.
5485 * @sds: sched_domain statistics
5486 * @sg: sched_group candidate to be checked for being the busiest
5487 * @sgs: sched_group statistics
5489 * Determine if @sg is a busier group than the previously selected
5492 * Return: %true if @sg is a busier group than the previously selected
5493 * busiest group. %false otherwise.
5495 static bool update_sd_pick_busiest(struct lb_env *env,
5496 struct sd_lb_stats *sds,
5497 struct sched_group *sg,
5498 struct sg_lb_stats *sgs)
5500 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5503 if (sgs->sum_nr_running > sgs->group_capacity)
5510 * ASYM_PACKING needs to move all the work to the lowest
5511 * numbered CPUs in the group, therefore mark all groups
5512 * higher than ourself as busy.
5514 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5515 env->dst_cpu < group_first_cpu(sg)) {
5519 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5526 #ifdef CONFIG_NUMA_BALANCING
5527 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5529 if (sgs->sum_nr_running > sgs->nr_numa_running)
5531 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5536 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5538 if (rq->nr_running > rq->nr_numa_running)
5540 if (rq->nr_running > rq->nr_preferred_running)
5545 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5550 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5554 #endif /* CONFIG_NUMA_BALANCING */
5557 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5558 * @env: The load balancing environment.
5559 * @balance: Should we balance.
5560 * @sds: variable to hold the statistics for this sched_domain.
5562 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5564 struct sched_domain *child = env->sd->child;
5565 struct sched_group *sg = env->sd->groups;
5566 struct sg_lb_stats tmp_sgs;
5567 int load_idx, prefer_sibling = 0;
5569 if (child && child->flags & SD_PREFER_SIBLING)
5572 load_idx = get_sd_load_idx(env->sd, env->idle);
5575 struct sg_lb_stats *sgs = &tmp_sgs;
5578 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5581 sgs = &sds->local_stat;
5583 if (env->idle != CPU_NEWLY_IDLE ||
5584 time_after_eq(jiffies, sg->sgp->next_update))
5585 update_group_power(env->sd, env->dst_cpu);
5588 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5594 * In case the child domain prefers tasks go to siblings
5595 * first, lower the sg capacity to one so that we'll try
5596 * and move all the excess tasks away. We lower the capacity
5597 * of a group only if the local group has the capacity to fit
5598 * these excess tasks, i.e. nr_running < group_capacity. The
5599 * extra check prevents the case where you always pull from the
5600 * heaviest group when it is already under-utilized (possible
5601 * with a large weight task outweighs the tasks on the system).
5603 if (prefer_sibling && sds->local &&
5604 sds->local_stat.group_has_capacity)
5605 sgs->group_capacity = min(sgs->group_capacity, 1U);
5607 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5609 sds->busiest_stat = *sgs;
5613 /* Now, start updating sd_lb_stats */
5614 sds->total_load += sgs->group_load;
5615 sds->total_pwr += sgs->group_power;
5618 } while (sg != env->sd->groups);
5620 if (env->sd->flags & SD_NUMA)
5621 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5625 * check_asym_packing - Check to see if the group is packed into the
5628 * This is primarily intended to used at the sibling level. Some
5629 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5630 * case of POWER7, it can move to lower SMT modes only when higher
5631 * threads are idle. When in lower SMT modes, the threads will
5632 * perform better since they share less core resources. Hence when we
5633 * have idle threads, we want them to be the higher ones.
5635 * This packing function is run on idle threads. It checks to see if
5636 * the busiest CPU in this domain (core in the P7 case) has a higher
5637 * CPU number than the packing function is being run on. Here we are
5638 * assuming lower CPU number will be equivalent to lower a SMT thread
5641 * Return: 1 when packing is required and a task should be moved to
5642 * this CPU. The amount of the imbalance is returned in *imbalance.
5644 * @env: The load balancing environment.
5645 * @sds: Statistics of the sched_domain which is to be packed
5647 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5651 if (!(env->sd->flags & SD_ASYM_PACKING))
5657 busiest_cpu = group_first_cpu(sds->busiest);
5658 if (env->dst_cpu > busiest_cpu)
5661 env->imbalance = DIV_ROUND_CLOSEST(
5662 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5669 * fix_small_imbalance - Calculate the minor imbalance that exists
5670 * amongst the groups of a sched_domain, during
5672 * @env: The load balancing environment.
5673 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5676 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5678 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5679 unsigned int imbn = 2;
5680 unsigned long scaled_busy_load_per_task;
5681 struct sg_lb_stats *local, *busiest;
5683 local = &sds->local_stat;
5684 busiest = &sds->busiest_stat;
5686 if (!local->sum_nr_running)
5687 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5688 else if (busiest->load_per_task > local->load_per_task)
5691 scaled_busy_load_per_task =
5692 (busiest->load_per_task * SCHED_POWER_SCALE) /
5693 busiest->group_power;
5695 if (busiest->avg_load + scaled_busy_load_per_task >=
5696 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5697 env->imbalance = busiest->load_per_task;
5702 * OK, we don't have enough imbalance to justify moving tasks,
5703 * however we may be able to increase total CPU power used by
5707 pwr_now += busiest->group_power *
5708 min(busiest->load_per_task, busiest->avg_load);
5709 pwr_now += local->group_power *
5710 min(local->load_per_task, local->avg_load);
5711 pwr_now /= SCHED_POWER_SCALE;
5713 /* Amount of load we'd subtract */
5714 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5715 busiest->group_power;
5716 if (busiest->avg_load > tmp) {
5717 pwr_move += busiest->group_power *
5718 min(busiest->load_per_task,
5719 busiest->avg_load - tmp);
5722 /* Amount of load we'd add */
5723 if (busiest->avg_load * busiest->group_power <
5724 busiest->load_per_task * SCHED_POWER_SCALE) {
5725 tmp = (busiest->avg_load * busiest->group_power) /
5728 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5731 pwr_move += local->group_power *
5732 min(local->load_per_task, local->avg_load + tmp);
5733 pwr_move /= SCHED_POWER_SCALE;
5735 /* Move if we gain throughput */
5736 if (pwr_move > pwr_now)
5737 env->imbalance = busiest->load_per_task;
5741 * calculate_imbalance - Calculate the amount of imbalance present within the
5742 * groups of a given sched_domain during load balance.
5743 * @env: load balance environment
5744 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5746 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5748 unsigned long max_pull, load_above_capacity = ~0UL;
5749 struct sg_lb_stats *local, *busiest;
5751 local = &sds->local_stat;
5752 busiest = &sds->busiest_stat;
5754 if (busiest->group_imb) {
5756 * In the group_imb case we cannot rely on group-wide averages
5757 * to ensure cpu-load equilibrium, look at wider averages. XXX
5759 busiest->load_per_task =
5760 min(busiest->load_per_task, sds->avg_load);
5764 * In the presence of smp nice balancing, certain scenarios can have
5765 * max load less than avg load(as we skip the groups at or below
5766 * its cpu_power, while calculating max_load..)
5768 if (busiest->avg_load <= sds->avg_load ||
5769 local->avg_load >= sds->avg_load) {
5771 return fix_small_imbalance(env, sds);
5774 if (!busiest->group_imb) {
5776 * Don't want to pull so many tasks that a group would go idle.
5777 * Except of course for the group_imb case, since then we might
5778 * have to drop below capacity to reach cpu-load equilibrium.
5780 load_above_capacity =
5781 (busiest->sum_nr_running - busiest->group_capacity);
5783 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5784 load_above_capacity /= busiest->group_power;
5788 * We're trying to get all the cpus to the average_load, so we don't
5789 * want to push ourselves above the average load, nor do we wish to
5790 * reduce the max loaded cpu below the average load. At the same time,
5791 * we also don't want to reduce the group load below the group capacity
5792 * (so that we can implement power-savings policies etc). Thus we look
5793 * for the minimum possible imbalance.
5795 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5797 /* How much load to actually move to equalise the imbalance */
5798 env->imbalance = min(
5799 max_pull * busiest->group_power,
5800 (sds->avg_load - local->avg_load) * local->group_power
5801 ) / SCHED_POWER_SCALE;
5804 * if *imbalance is less than the average load per runnable task
5805 * there is no guarantee that any tasks will be moved so we'll have
5806 * a think about bumping its value to force at least one task to be
5809 if (env->imbalance < busiest->load_per_task)
5810 return fix_small_imbalance(env, sds);
5813 /******* find_busiest_group() helpers end here *********************/
5816 * find_busiest_group - Returns the busiest group within the sched_domain
5817 * if there is an imbalance. If there isn't an imbalance, and
5818 * the user has opted for power-savings, it returns a group whose
5819 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5820 * such a group exists.
5822 * Also calculates the amount of weighted load which should be moved
5823 * to restore balance.
5825 * @env: The load balancing environment.
5827 * Return: - The busiest group if imbalance exists.
5828 * - If no imbalance and user has opted for power-savings balance,
5829 * return the least loaded group whose CPUs can be
5830 * put to idle by rebalancing its tasks onto our group.
5832 static struct sched_group *find_busiest_group(struct lb_env *env)
5834 struct sg_lb_stats *local, *busiest;
5835 struct sd_lb_stats sds;
5837 init_sd_lb_stats(&sds);
5840 * Compute the various statistics relavent for load balancing at
5843 update_sd_lb_stats(env, &sds);
5844 local = &sds.local_stat;
5845 busiest = &sds.busiest_stat;
5847 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5848 check_asym_packing(env, &sds))
5851 /* There is no busy sibling group to pull tasks from */
5852 if (!sds.busiest || busiest->sum_nr_running == 0)
5855 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5858 * If the busiest group is imbalanced the below checks don't
5859 * work because they assume all things are equal, which typically
5860 * isn't true due to cpus_allowed constraints and the like.
5862 if (busiest->group_imb)
5865 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5866 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5867 !busiest->group_has_capacity)
5871 * If the local group is more busy than the selected busiest group
5872 * don't try and pull any tasks.
5874 if (local->avg_load >= busiest->avg_load)
5878 * Don't pull any tasks if this group is already above the domain
5881 if (local->avg_load >= sds.avg_load)
5884 if (env->idle == CPU_IDLE) {
5886 * This cpu is idle. If the busiest group load doesn't
5887 * have more tasks than the number of available cpu's and
5888 * there is no imbalance between this and busiest group
5889 * wrt to idle cpu's, it is balanced.
5891 if ((local->idle_cpus < busiest->idle_cpus) &&
5892 busiest->sum_nr_running <= busiest->group_weight)
5896 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5897 * imbalance_pct to be conservative.
5899 if (100 * busiest->avg_load <=
5900 env->sd->imbalance_pct * local->avg_load)
5905 /* Looks like there is an imbalance. Compute it */
5906 calculate_imbalance(env, &sds);
5915 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5917 static struct rq *find_busiest_queue(struct lb_env *env,
5918 struct sched_group *group)
5920 struct rq *busiest = NULL, *rq;
5921 unsigned long busiest_load = 0, busiest_power = 1;
5924 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5925 unsigned long power, capacity, wl;
5929 rt = fbq_classify_rq(rq);
5932 * We classify groups/runqueues into three groups:
5933 * - regular: there are !numa tasks
5934 * - remote: there are numa tasks that run on the 'wrong' node
5935 * - all: there is no distinction
5937 * In order to avoid migrating ideally placed numa tasks,
5938 * ignore those when there's better options.
5940 * If we ignore the actual busiest queue to migrate another
5941 * task, the next balance pass can still reduce the busiest
5942 * queue by moving tasks around inside the node.
5944 * If we cannot move enough load due to this classification
5945 * the next pass will adjust the group classification and
5946 * allow migration of more tasks.
5948 * Both cases only affect the total convergence complexity.
5950 if (rt > env->fbq_type)
5953 power = power_of(i);
5954 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5956 capacity = fix_small_capacity(env->sd, group);
5958 wl = weighted_cpuload(i);
5961 * When comparing with imbalance, use weighted_cpuload()
5962 * which is not scaled with the cpu power.
5964 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5968 * For the load comparisons with the other cpu's, consider
5969 * the weighted_cpuload() scaled with the cpu power, so that
5970 * the load can be moved away from the cpu that is potentially
5971 * running at a lower capacity.
5973 * Thus we're looking for max(wl_i / power_i), crosswise
5974 * multiplication to rid ourselves of the division works out
5975 * to: wl_i * power_j > wl_j * power_i; where j is our
5978 if (wl * busiest_power > busiest_load * power) {
5980 busiest_power = power;
5989 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5990 * so long as it is large enough.
5992 #define MAX_PINNED_INTERVAL 512
5994 /* Working cpumask for load_balance and load_balance_newidle. */
5995 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5997 static int need_active_balance(struct lb_env *env)
5999 struct sched_domain *sd = env->sd;
6001 if (env->idle == CPU_NEWLY_IDLE) {
6004 * ASYM_PACKING needs to force migrate tasks from busy but
6005 * higher numbered CPUs in order to pack all tasks in the
6006 * lowest numbered CPUs.
6008 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6012 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6015 static int active_load_balance_cpu_stop(void *data);
6017 static int should_we_balance(struct lb_env *env)
6019 struct sched_group *sg = env->sd->groups;
6020 struct cpumask *sg_cpus, *sg_mask;
6021 int cpu, balance_cpu = -1;
6024 * In the newly idle case, we will allow all the cpu's
6025 * to do the newly idle load balance.
6027 if (env->idle == CPU_NEWLY_IDLE)
6030 sg_cpus = sched_group_cpus(sg);
6031 sg_mask = sched_group_mask(sg);
6032 /* Try to find first idle cpu */
6033 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6034 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6041 if (balance_cpu == -1)
6042 balance_cpu = group_balance_cpu(sg);
6045 * First idle cpu or the first cpu(busiest) in this sched group
6046 * is eligible for doing load balancing at this and above domains.
6048 return balance_cpu == env->dst_cpu;
6052 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6053 * tasks if there is an imbalance.
6055 static int load_balance(int this_cpu, struct rq *this_rq,
6056 struct sched_domain *sd, enum cpu_idle_type idle,
6057 int *continue_balancing)
6059 int ld_moved, cur_ld_moved, active_balance = 0;
6060 struct sched_domain *sd_parent = sd->parent;
6061 struct sched_group *group;
6063 unsigned long flags;
6064 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6066 struct lb_env env = {
6068 .dst_cpu = this_cpu,
6070 .dst_grpmask = sched_group_cpus(sd->groups),
6072 .loop_break = sched_nr_migrate_break,
6078 * For NEWLY_IDLE load_balancing, we don't need to consider
6079 * other cpus in our group
6081 if (idle == CPU_NEWLY_IDLE)
6082 env.dst_grpmask = NULL;
6084 cpumask_copy(cpus, cpu_active_mask);
6086 schedstat_inc(sd, lb_count[idle]);
6089 if (!should_we_balance(&env)) {
6090 *continue_balancing = 0;
6094 group = find_busiest_group(&env);
6096 schedstat_inc(sd, lb_nobusyg[idle]);
6100 busiest = find_busiest_queue(&env, group);
6102 schedstat_inc(sd, lb_nobusyq[idle]);
6106 BUG_ON(busiest == env.dst_rq);
6108 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6111 if (busiest->nr_running > 1) {
6113 * Attempt to move tasks. If find_busiest_group has found
6114 * an imbalance but busiest->nr_running <= 1, the group is
6115 * still unbalanced. ld_moved simply stays zero, so it is
6116 * correctly treated as an imbalance.
6118 env.flags |= LBF_ALL_PINNED;
6119 env.src_cpu = busiest->cpu;
6120 env.src_rq = busiest;
6121 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6124 local_irq_save(flags);
6125 double_rq_lock(env.dst_rq, busiest);
6128 * cur_ld_moved - load moved in current iteration
6129 * ld_moved - cumulative load moved across iterations
6131 cur_ld_moved = move_tasks(&env);
6132 ld_moved += cur_ld_moved;
6133 double_rq_unlock(env.dst_rq, busiest);
6134 local_irq_restore(flags);
6137 * some other cpu did the load balance for us.
6139 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6140 resched_cpu(env.dst_cpu);
6142 if (env.flags & LBF_NEED_BREAK) {
6143 env.flags &= ~LBF_NEED_BREAK;
6148 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6149 * us and move them to an alternate dst_cpu in our sched_group
6150 * where they can run. The upper limit on how many times we
6151 * iterate on same src_cpu is dependent on number of cpus in our
6154 * This changes load balance semantics a bit on who can move
6155 * load to a given_cpu. In addition to the given_cpu itself
6156 * (or a ilb_cpu acting on its behalf where given_cpu is
6157 * nohz-idle), we now have balance_cpu in a position to move
6158 * load to given_cpu. In rare situations, this may cause
6159 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6160 * _independently_ and at _same_ time to move some load to
6161 * given_cpu) causing exceess load to be moved to given_cpu.
6162 * This however should not happen so much in practice and
6163 * moreover subsequent load balance cycles should correct the
6164 * excess load moved.
6166 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6168 /* Prevent to re-select dst_cpu via env's cpus */
6169 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6171 env.dst_rq = cpu_rq(env.new_dst_cpu);
6172 env.dst_cpu = env.new_dst_cpu;
6173 env.flags &= ~LBF_DST_PINNED;
6175 env.loop_break = sched_nr_migrate_break;
6178 * Go back to "more_balance" rather than "redo" since we
6179 * need to continue with same src_cpu.
6185 * We failed to reach balance because of affinity.
6188 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6190 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6191 *group_imbalance = 1;
6192 } else if (*group_imbalance)
6193 *group_imbalance = 0;
6196 /* All tasks on this runqueue were pinned by CPU affinity */
6197 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6198 cpumask_clear_cpu(cpu_of(busiest), cpus);
6199 if (!cpumask_empty(cpus)) {
6201 env.loop_break = sched_nr_migrate_break;
6209 schedstat_inc(sd, lb_failed[idle]);
6211 * Increment the failure counter only on periodic balance.
6212 * We do not want newidle balance, which can be very
6213 * frequent, pollute the failure counter causing
6214 * excessive cache_hot migrations and active balances.
6216 if (idle != CPU_NEWLY_IDLE)
6217 sd->nr_balance_failed++;
6219 if (need_active_balance(&env)) {
6220 raw_spin_lock_irqsave(&busiest->lock, flags);
6222 /* don't kick the active_load_balance_cpu_stop,
6223 * if the curr task on busiest cpu can't be
6226 if (!cpumask_test_cpu(this_cpu,
6227 tsk_cpus_allowed(busiest->curr))) {
6228 raw_spin_unlock_irqrestore(&busiest->lock,
6230 env.flags |= LBF_ALL_PINNED;
6231 goto out_one_pinned;
6235 * ->active_balance synchronizes accesses to
6236 * ->active_balance_work. Once set, it's cleared
6237 * only after active load balance is finished.
6239 if (!busiest->active_balance) {
6240 busiest->active_balance = 1;
6241 busiest->push_cpu = this_cpu;
6244 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6246 if (active_balance) {
6247 stop_one_cpu_nowait(cpu_of(busiest),
6248 active_load_balance_cpu_stop, busiest,
6249 &busiest->active_balance_work);
6253 * We've kicked active balancing, reset the failure
6256 sd->nr_balance_failed = sd->cache_nice_tries+1;
6259 sd->nr_balance_failed = 0;
6261 if (likely(!active_balance)) {
6262 /* We were unbalanced, so reset the balancing interval */
6263 sd->balance_interval = sd->min_interval;
6266 * If we've begun active balancing, start to back off. This
6267 * case may not be covered by the all_pinned logic if there
6268 * is only 1 task on the busy runqueue (because we don't call
6271 if (sd->balance_interval < sd->max_interval)
6272 sd->balance_interval *= 2;
6278 schedstat_inc(sd, lb_balanced[idle]);
6280 sd->nr_balance_failed = 0;
6283 /* tune up the balancing interval */
6284 if (((env.flags & LBF_ALL_PINNED) &&
6285 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6286 (sd->balance_interval < sd->max_interval))
6287 sd->balance_interval *= 2;
6295 * idle_balance is called by schedule() if this_cpu is about to become
6296 * idle. Attempts to pull tasks from other CPUs.
6298 void idle_balance(int this_cpu, struct rq *this_rq)
6300 struct sched_domain *sd;
6301 int pulled_task = 0;
6302 unsigned long next_balance = jiffies + HZ;
6305 this_rq->idle_stamp = rq_clock(this_rq);
6307 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6311 * Drop the rq->lock, but keep IRQ/preempt disabled.
6313 raw_spin_unlock(&this_rq->lock);
6315 update_blocked_averages(this_cpu);
6317 for_each_domain(this_cpu, sd) {
6318 unsigned long interval;
6319 int continue_balancing = 1;
6320 u64 t0, domain_cost;
6322 if (!(sd->flags & SD_LOAD_BALANCE))
6325 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6328 if (sd->flags & SD_BALANCE_NEWIDLE) {
6329 t0 = sched_clock_cpu(this_cpu);
6331 /* If we've pulled tasks over stop searching: */
6332 pulled_task = load_balance(this_cpu, this_rq,
6334 &continue_balancing);
6336 domain_cost = sched_clock_cpu(this_cpu) - t0;
6337 if (domain_cost > sd->max_newidle_lb_cost)
6338 sd->max_newidle_lb_cost = domain_cost;
6340 curr_cost += domain_cost;
6343 interval = msecs_to_jiffies(sd->balance_interval);
6344 if (time_after(next_balance, sd->last_balance + interval))
6345 next_balance = sd->last_balance + interval;
6347 this_rq->idle_stamp = 0;
6353 raw_spin_lock(&this_rq->lock);
6355 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6357 * We are going idle. next_balance may be set based on
6358 * a busy processor. So reset next_balance.
6360 this_rq->next_balance = next_balance;
6363 if (curr_cost > this_rq->max_idle_balance_cost)
6364 this_rq->max_idle_balance_cost = curr_cost;
6368 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6369 * running tasks off the busiest CPU onto idle CPUs. It requires at
6370 * least 1 task to be running on each physical CPU where possible, and
6371 * avoids physical / logical imbalances.
6373 static int active_load_balance_cpu_stop(void *data)
6375 struct rq *busiest_rq = data;
6376 int busiest_cpu = cpu_of(busiest_rq);
6377 int target_cpu = busiest_rq->push_cpu;
6378 struct rq *target_rq = cpu_rq(target_cpu);
6379 struct sched_domain *sd;
6381 raw_spin_lock_irq(&busiest_rq->lock);
6383 /* make sure the requested cpu hasn't gone down in the meantime */
6384 if (unlikely(busiest_cpu != smp_processor_id() ||
6385 !busiest_rq->active_balance))
6388 /* Is there any task to move? */
6389 if (busiest_rq->nr_running <= 1)
6393 * This condition is "impossible", if it occurs
6394 * we need to fix it. Originally reported by
6395 * Bjorn Helgaas on a 128-cpu setup.
6397 BUG_ON(busiest_rq == target_rq);
6399 /* move a task from busiest_rq to target_rq */
6400 double_lock_balance(busiest_rq, target_rq);
6402 /* Search for an sd spanning us and the target CPU. */
6404 for_each_domain(target_cpu, sd) {
6405 if ((sd->flags & SD_LOAD_BALANCE) &&
6406 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6411 struct lb_env env = {
6413 .dst_cpu = target_cpu,
6414 .dst_rq = target_rq,
6415 .src_cpu = busiest_rq->cpu,
6416 .src_rq = busiest_rq,
6420 schedstat_inc(sd, alb_count);
6422 if (move_one_task(&env))
6423 schedstat_inc(sd, alb_pushed);
6425 schedstat_inc(sd, alb_failed);
6428 double_unlock_balance(busiest_rq, target_rq);
6430 busiest_rq->active_balance = 0;
6431 raw_spin_unlock_irq(&busiest_rq->lock);
6435 #ifdef CONFIG_NO_HZ_COMMON
6437 * idle load balancing details
6438 * - When one of the busy CPUs notice that there may be an idle rebalancing
6439 * needed, they will kick the idle load balancer, which then does idle
6440 * load balancing for all the idle CPUs.
6443 cpumask_var_t idle_cpus_mask;
6445 unsigned long next_balance; /* in jiffy units */
6446 } nohz ____cacheline_aligned;
6448 static inline int find_new_ilb(int call_cpu)
6450 int ilb = cpumask_first(nohz.idle_cpus_mask);
6452 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6459 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6460 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6461 * CPU (if there is one).
6463 static void nohz_balancer_kick(int cpu)
6467 nohz.next_balance++;
6469 ilb_cpu = find_new_ilb(cpu);
6471 if (ilb_cpu >= nr_cpu_ids)
6474 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6477 * Use smp_send_reschedule() instead of resched_cpu().
6478 * This way we generate a sched IPI on the target cpu which
6479 * is idle. And the softirq performing nohz idle load balance
6480 * will be run before returning from the IPI.
6482 smp_send_reschedule(ilb_cpu);
6486 static inline void nohz_balance_exit_idle(int cpu)
6488 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6489 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6490 atomic_dec(&nohz.nr_cpus);
6491 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6495 static inline void set_cpu_sd_state_busy(void)
6497 struct sched_domain *sd;
6500 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6502 if (!sd || !sd->nohz_idle)
6506 for (; sd; sd = sd->parent)
6507 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6512 void set_cpu_sd_state_idle(void)
6514 struct sched_domain *sd;
6517 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6519 if (!sd || sd->nohz_idle)
6523 for (; sd; sd = sd->parent)
6524 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6530 * This routine will record that the cpu is going idle with tick stopped.
6531 * This info will be used in performing idle load balancing in the future.
6533 void nohz_balance_enter_idle(int cpu)
6536 * If this cpu is going down, then nothing needs to be done.
6538 if (!cpu_active(cpu))
6541 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6544 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6545 atomic_inc(&nohz.nr_cpus);
6546 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6549 static int sched_ilb_notifier(struct notifier_block *nfb,
6550 unsigned long action, void *hcpu)
6552 switch (action & ~CPU_TASKS_FROZEN) {
6554 nohz_balance_exit_idle(smp_processor_id());
6562 static DEFINE_SPINLOCK(balancing);
6565 * Scale the max load_balance interval with the number of CPUs in the system.
6566 * This trades load-balance latency on larger machines for less cross talk.
6568 void update_max_interval(void)
6570 max_load_balance_interval = HZ*num_online_cpus()/10;
6574 * It checks each scheduling domain to see if it is due to be balanced,
6575 * and initiates a balancing operation if so.
6577 * Balancing parameters are set up in init_sched_domains.
6579 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6581 int continue_balancing = 1;
6582 struct rq *rq = cpu_rq(cpu);
6583 unsigned long interval;
6584 struct sched_domain *sd;
6585 /* Earliest time when we have to do rebalance again */
6586 unsigned long next_balance = jiffies + 60*HZ;
6587 int update_next_balance = 0;
6588 int need_serialize, need_decay = 0;
6591 update_blocked_averages(cpu);
6594 for_each_domain(cpu, sd) {
6596 * Decay the newidle max times here because this is a regular
6597 * visit to all the domains. Decay ~1% per second.
6599 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6600 sd->max_newidle_lb_cost =
6601 (sd->max_newidle_lb_cost * 253) / 256;
6602 sd->next_decay_max_lb_cost = jiffies + HZ;
6605 max_cost += sd->max_newidle_lb_cost;
6607 if (!(sd->flags & SD_LOAD_BALANCE))
6611 * Stop the load balance at this level. There is another
6612 * CPU in our sched group which is doing load balancing more
6615 if (!continue_balancing) {
6621 interval = sd->balance_interval;
6622 if (idle != CPU_IDLE)
6623 interval *= sd->busy_factor;
6625 /* scale ms to jiffies */
6626 interval = msecs_to_jiffies(interval);
6627 interval = clamp(interval, 1UL, max_load_balance_interval);
6629 need_serialize = sd->flags & SD_SERIALIZE;
6631 if (need_serialize) {
6632 if (!spin_trylock(&balancing))
6636 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6637 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6639 * The LBF_DST_PINNED logic could have changed
6640 * env->dst_cpu, so we can't know our idle
6641 * state even if we migrated tasks. Update it.
6643 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6645 sd->last_balance = jiffies;
6648 spin_unlock(&balancing);
6650 if (time_after(next_balance, sd->last_balance + interval)) {
6651 next_balance = sd->last_balance + interval;
6652 update_next_balance = 1;
6657 * Ensure the rq-wide value also decays but keep it at a
6658 * reasonable floor to avoid funnies with rq->avg_idle.
6660 rq->max_idle_balance_cost =
6661 max((u64)sysctl_sched_migration_cost, max_cost);
6666 * next_balance will be updated only when there is a need.
6667 * When the cpu is attached to null domain for ex, it will not be
6670 if (likely(update_next_balance))
6671 rq->next_balance = next_balance;
6674 #ifdef CONFIG_NO_HZ_COMMON
6676 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6677 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6679 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6681 struct rq *this_rq = cpu_rq(this_cpu);
6685 if (idle != CPU_IDLE ||
6686 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6689 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6690 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6694 * If this cpu gets work to do, stop the load balancing
6695 * work being done for other cpus. Next load
6696 * balancing owner will pick it up.
6701 rq = cpu_rq(balance_cpu);
6703 raw_spin_lock_irq(&rq->lock);
6704 update_rq_clock(rq);
6705 update_idle_cpu_load(rq);
6706 raw_spin_unlock_irq(&rq->lock);
6708 rebalance_domains(balance_cpu, CPU_IDLE);
6710 if (time_after(this_rq->next_balance, rq->next_balance))
6711 this_rq->next_balance = rq->next_balance;
6713 nohz.next_balance = this_rq->next_balance;
6715 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6719 * Current heuristic for kicking the idle load balancer in the presence
6720 * of an idle cpu is the system.
6721 * - This rq has more than one task.
6722 * - At any scheduler domain level, this cpu's scheduler group has multiple
6723 * busy cpu's exceeding the group's power.
6724 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6725 * domain span are idle.
6727 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6729 unsigned long now = jiffies;
6730 struct sched_domain *sd;
6732 if (unlikely(idle_cpu(cpu)))
6736 * We may be recently in ticked or tickless idle mode. At the first
6737 * busy tick after returning from idle, we will update the busy stats.
6739 set_cpu_sd_state_busy();
6740 nohz_balance_exit_idle(cpu);
6743 * None are in tickless mode and hence no need for NOHZ idle load
6746 if (likely(!atomic_read(&nohz.nr_cpus)))
6749 if (time_before(now, nohz.next_balance))
6752 if (rq->nr_running >= 2)
6756 for_each_domain(cpu, sd) {
6757 struct sched_group *sg = sd->groups;
6758 struct sched_group_power *sgp = sg->sgp;
6759 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6761 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6762 goto need_kick_unlock;
6764 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6765 && (cpumask_first_and(nohz.idle_cpus_mask,
6766 sched_domain_span(sd)) < cpu))
6767 goto need_kick_unlock;
6769 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6781 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6785 * run_rebalance_domains is triggered when needed from the scheduler tick.
6786 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6788 static void run_rebalance_domains(struct softirq_action *h)
6790 int this_cpu = smp_processor_id();
6791 struct rq *this_rq = cpu_rq(this_cpu);
6792 enum cpu_idle_type idle = this_rq->idle_balance ?
6793 CPU_IDLE : CPU_NOT_IDLE;
6795 rebalance_domains(this_cpu, idle);
6798 * If this cpu has a pending nohz_balance_kick, then do the
6799 * balancing on behalf of the other idle cpus whose ticks are
6802 nohz_idle_balance(this_cpu, idle);
6805 static inline int on_null_domain(int cpu)
6807 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6811 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6813 void trigger_load_balance(struct rq *rq, int cpu)
6815 /* Don't need to rebalance while attached to NULL domain */
6816 if (time_after_eq(jiffies, rq->next_balance) &&
6817 likely(!on_null_domain(cpu)))
6818 raise_softirq(SCHED_SOFTIRQ);
6819 #ifdef CONFIG_NO_HZ_COMMON
6820 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6821 nohz_balancer_kick(cpu);
6825 static void rq_online_fair(struct rq *rq)
6830 static void rq_offline_fair(struct rq *rq)
6834 /* Ensure any throttled groups are reachable by pick_next_task */
6835 unthrottle_offline_cfs_rqs(rq);
6838 #endif /* CONFIG_SMP */
6841 * scheduler tick hitting a task of our scheduling class:
6843 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6845 struct cfs_rq *cfs_rq;
6846 struct sched_entity *se = &curr->se;
6848 for_each_sched_entity(se) {
6849 cfs_rq = cfs_rq_of(se);
6850 entity_tick(cfs_rq, se, queued);
6853 if (numabalancing_enabled)
6854 task_tick_numa(rq, curr);
6856 update_rq_runnable_avg(rq, 1);
6860 * called on fork with the child task as argument from the parent's context
6861 * - child not yet on the tasklist
6862 * - preemption disabled
6864 static void task_fork_fair(struct task_struct *p)
6866 struct cfs_rq *cfs_rq;
6867 struct sched_entity *se = &p->se, *curr;
6868 int this_cpu = smp_processor_id();
6869 struct rq *rq = this_rq();
6870 unsigned long flags;
6872 raw_spin_lock_irqsave(&rq->lock, flags);
6874 update_rq_clock(rq);
6876 cfs_rq = task_cfs_rq(current);
6877 curr = cfs_rq->curr;
6880 * Not only the cpu but also the task_group of the parent might have
6881 * been changed after parent->se.parent,cfs_rq were copied to
6882 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6883 * of child point to valid ones.
6886 __set_task_cpu(p, this_cpu);
6889 update_curr(cfs_rq);
6892 se->vruntime = curr->vruntime;
6893 place_entity(cfs_rq, se, 1);
6895 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6897 * Upon rescheduling, sched_class::put_prev_task() will place
6898 * 'current' within the tree based on its new key value.
6900 swap(curr->vruntime, se->vruntime);
6901 resched_task(rq->curr);
6904 se->vruntime -= cfs_rq->min_vruntime;
6906 raw_spin_unlock_irqrestore(&rq->lock, flags);
6910 * Priority of the task has changed. Check to see if we preempt
6914 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6920 * Reschedule if we are currently running on this runqueue and
6921 * our priority decreased, or if we are not currently running on
6922 * this runqueue and our priority is higher than the current's
6924 if (rq->curr == p) {
6925 if (p->prio > oldprio)
6926 resched_task(rq->curr);
6928 check_preempt_curr(rq, p, 0);
6931 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6933 struct sched_entity *se = &p->se;
6934 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6937 * Ensure the task's vruntime is normalized, so that when its
6938 * switched back to the fair class the enqueue_entity(.flags=0) will
6939 * do the right thing.
6941 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6942 * have normalized the vruntime, if it was !on_rq, then only when
6943 * the task is sleeping will it still have non-normalized vruntime.
6945 if (!se->on_rq && p->state != TASK_RUNNING) {
6947 * Fix up our vruntime so that the current sleep doesn't
6948 * cause 'unlimited' sleep bonus.
6950 place_entity(cfs_rq, se, 0);
6951 se->vruntime -= cfs_rq->min_vruntime;
6956 * Remove our load from contribution when we leave sched_fair
6957 * and ensure we don't carry in an old decay_count if we
6960 if (se->avg.decay_count) {
6961 __synchronize_entity_decay(se);
6962 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6968 * We switched to the sched_fair class.
6970 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6976 * We were most likely switched from sched_rt, so
6977 * kick off the schedule if running, otherwise just see
6978 * if we can still preempt the current task.
6981 resched_task(rq->curr);
6983 check_preempt_curr(rq, p, 0);
6986 /* Account for a task changing its policy or group.
6988 * This routine is mostly called to set cfs_rq->curr field when a task
6989 * migrates between groups/classes.
6991 static void set_curr_task_fair(struct rq *rq)
6993 struct sched_entity *se = &rq->curr->se;
6995 for_each_sched_entity(se) {
6996 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6998 set_next_entity(cfs_rq, se);
6999 /* ensure bandwidth has been allocated on our new cfs_rq */
7000 account_cfs_rq_runtime(cfs_rq, 0);
7004 void init_cfs_rq(struct cfs_rq *cfs_rq)
7006 cfs_rq->tasks_timeline = RB_ROOT;
7007 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7008 #ifndef CONFIG_64BIT
7009 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7012 atomic64_set(&cfs_rq->decay_counter, 1);
7013 atomic_long_set(&cfs_rq->removed_load, 0);
7017 #ifdef CONFIG_FAIR_GROUP_SCHED
7018 static void task_move_group_fair(struct task_struct *p, int on_rq)
7020 struct cfs_rq *cfs_rq;
7022 * If the task was not on the rq at the time of this cgroup movement
7023 * it must have been asleep, sleeping tasks keep their ->vruntime
7024 * absolute on their old rq until wakeup (needed for the fair sleeper
7025 * bonus in place_entity()).
7027 * If it was on the rq, we've just 'preempted' it, which does convert
7028 * ->vruntime to a relative base.
7030 * Make sure both cases convert their relative position when migrating
7031 * to another cgroup's rq. This does somewhat interfere with the
7032 * fair sleeper stuff for the first placement, but who cares.
7035 * When !on_rq, vruntime of the task has usually NOT been normalized.
7036 * But there are some cases where it has already been normalized:
7038 * - Moving a forked child which is waiting for being woken up by
7039 * wake_up_new_task().
7040 * - Moving a task which has been woken up by try_to_wake_up() and
7041 * waiting for actually being woken up by sched_ttwu_pending().
7043 * To prevent boost or penalty in the new cfs_rq caused by delta
7044 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7046 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7050 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7051 set_task_rq(p, task_cpu(p));
7053 cfs_rq = cfs_rq_of(&p->se);
7054 p->se.vruntime += cfs_rq->min_vruntime;
7057 * migrate_task_rq_fair() will have removed our previous
7058 * contribution, but we must synchronize for ongoing future
7061 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7062 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7067 void free_fair_sched_group(struct task_group *tg)
7071 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7073 for_each_possible_cpu(i) {
7075 kfree(tg->cfs_rq[i]);
7084 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7086 struct cfs_rq *cfs_rq;
7087 struct sched_entity *se;
7090 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7093 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7097 tg->shares = NICE_0_LOAD;
7099 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7101 for_each_possible_cpu(i) {
7102 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7103 GFP_KERNEL, cpu_to_node(i));
7107 se = kzalloc_node(sizeof(struct sched_entity),
7108 GFP_KERNEL, cpu_to_node(i));
7112 init_cfs_rq(cfs_rq);
7113 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7124 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7126 struct rq *rq = cpu_rq(cpu);
7127 unsigned long flags;
7130 * Only empty task groups can be destroyed; so we can speculatively
7131 * check on_list without danger of it being re-added.
7133 if (!tg->cfs_rq[cpu]->on_list)
7136 raw_spin_lock_irqsave(&rq->lock, flags);
7137 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7138 raw_spin_unlock_irqrestore(&rq->lock, flags);
7141 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7142 struct sched_entity *se, int cpu,
7143 struct sched_entity *parent)
7145 struct rq *rq = cpu_rq(cpu);
7149 init_cfs_rq_runtime(cfs_rq);
7151 tg->cfs_rq[cpu] = cfs_rq;
7154 /* se could be NULL for root_task_group */
7159 se->cfs_rq = &rq->cfs;
7161 se->cfs_rq = parent->my_q;
7164 update_load_set(&se->load, 0);
7165 se->parent = parent;
7168 static DEFINE_MUTEX(shares_mutex);
7170 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7173 unsigned long flags;
7176 * We can't change the weight of the root cgroup.
7181 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7183 mutex_lock(&shares_mutex);
7184 if (tg->shares == shares)
7187 tg->shares = shares;
7188 for_each_possible_cpu(i) {
7189 struct rq *rq = cpu_rq(i);
7190 struct sched_entity *se;
7193 /* Propagate contribution to hierarchy */
7194 raw_spin_lock_irqsave(&rq->lock, flags);
7196 /* Possible calls to update_curr() need rq clock */
7197 update_rq_clock(rq);
7198 for_each_sched_entity(se)
7199 update_cfs_shares(group_cfs_rq(se));
7200 raw_spin_unlock_irqrestore(&rq->lock, flags);
7204 mutex_unlock(&shares_mutex);
7207 #else /* CONFIG_FAIR_GROUP_SCHED */
7209 void free_fair_sched_group(struct task_group *tg) { }
7211 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7216 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7218 #endif /* CONFIG_FAIR_GROUP_SCHED */
7221 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7223 struct sched_entity *se = &task->se;
7224 unsigned int rr_interval = 0;
7227 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7230 if (rq->cfs.load.weight)
7231 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7237 * All the scheduling class methods:
7239 const struct sched_class fair_sched_class = {
7240 .next = &idle_sched_class,
7241 .enqueue_task = enqueue_task_fair,
7242 .dequeue_task = dequeue_task_fair,
7243 .yield_task = yield_task_fair,
7244 .yield_to_task = yield_to_task_fair,
7246 .check_preempt_curr = check_preempt_wakeup,
7248 .pick_next_task = pick_next_task_fair,
7249 .put_prev_task = put_prev_task_fair,
7252 .select_task_rq = select_task_rq_fair,
7253 .migrate_task_rq = migrate_task_rq_fair,
7255 .rq_online = rq_online_fair,
7256 .rq_offline = rq_offline_fair,
7258 .task_waking = task_waking_fair,
7261 .set_curr_task = set_curr_task_fair,
7262 .task_tick = task_tick_fair,
7263 .task_fork = task_fork_fair,
7265 .prio_changed = prio_changed_fair,
7266 .switched_from = switched_from_fair,
7267 .switched_to = switched_to_fair,
7269 .get_rr_interval = get_rr_interval_fair,
7271 #ifdef CONFIG_FAIR_GROUP_SCHED
7272 .task_move_group = task_move_group_fair,
7276 #ifdef CONFIG_SCHED_DEBUG
7277 void print_cfs_stats(struct seq_file *m, int cpu)
7279 struct cfs_rq *cfs_rq;
7282 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7283 print_cfs_rq(m, cpu, cfs_rq);
7288 __init void init_sched_fair_class(void)
7291 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7293 #ifdef CONFIG_NO_HZ_COMMON
7294 nohz.next_balance = jiffies;
7295 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7296 cpu_notifier(sched_ilb_notifier, 0);