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;
894 spinlock_t lock; /* nr_tasks, tasks */
896 struct list_head task_list;
899 atomic_long_t faults[0];
902 static inline int task_faults_idx(int nid, int priv)
904 return 2 * nid + priv;
907 static inline unsigned long task_faults(struct task_struct *p, int nid)
912 return p->numa_faults[task_faults_idx(nid, 0)] +
913 p->numa_faults[task_faults_idx(nid, 1)];
916 static unsigned long weighted_cpuload(const int cpu);
917 static unsigned long source_load(int cpu, int type);
918 static unsigned long target_load(int cpu, int type);
919 static unsigned long power_of(int cpu);
920 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
922 /* Cached statistics for all CPUs within a node */
924 unsigned long nr_running;
927 /* Total compute capacity of CPUs on a node */
930 /* Approximate capacity in terms of runnable tasks on a node */
931 unsigned long capacity;
936 * XXX borrowed from update_sg_lb_stats
938 static void update_numa_stats(struct numa_stats *ns, int nid)
942 memset(ns, 0, sizeof(*ns));
943 for_each_cpu(cpu, cpumask_of_node(nid)) {
944 struct rq *rq = cpu_rq(cpu);
946 ns->nr_running += rq->nr_running;
947 ns->load += weighted_cpuload(cpu);
948 ns->power += power_of(cpu);
951 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
952 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
953 ns->has_capacity = (ns->nr_running < ns->capacity);
956 struct task_numa_env {
957 struct task_struct *p;
959 int src_cpu, src_nid;
960 int dst_cpu, dst_nid;
962 struct numa_stats src_stats, dst_stats;
964 int imbalance_pct, idx;
966 struct task_struct *best_task;
971 static void task_numa_assign(struct task_numa_env *env,
972 struct task_struct *p, long imp)
975 put_task_struct(env->best_task);
981 env->best_cpu = env->dst_cpu;
985 * This checks if the overall compute and NUMA accesses of the system would
986 * be improved if the source tasks was migrated to the target dst_cpu taking
987 * into account that it might be best if task running on the dst_cpu should
988 * be exchanged with the source task
990 static void task_numa_compare(struct task_numa_env *env, long imp)
992 struct rq *src_rq = cpu_rq(env->src_cpu);
993 struct rq *dst_rq = cpu_rq(env->dst_cpu);
994 struct task_struct *cur;
995 long dst_load, src_load;
999 cur = ACCESS_ONCE(dst_rq->curr);
1000 if (cur->pid == 0) /* idle */
1004 * "imp" is the fault differential for the source task between the
1005 * source and destination node. Calculate the total differential for
1006 * the source task and potential destination task. The more negative
1007 * the value is, the more rmeote accesses that would be expected to
1008 * be incurred if the tasks were swapped.
1011 /* Skip this swap candidate if cannot move to the source cpu */
1012 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1015 imp += task_faults(cur, env->src_nid) -
1016 task_faults(cur, env->dst_nid);
1019 if (imp < env->best_imp)
1023 /* Is there capacity at our destination? */
1024 if (env->src_stats.has_capacity &&
1025 !env->dst_stats.has_capacity)
1031 /* Balance doesn't matter much if we're running a task per cpu */
1032 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1036 * In the overloaded case, try and keep the load balanced.
1039 dst_load = env->dst_stats.load;
1040 src_load = env->src_stats.load;
1042 /* XXX missing power terms */
1043 load = task_h_load(env->p);
1048 load = task_h_load(cur);
1053 /* make src_load the smaller */
1054 if (dst_load < src_load)
1055 swap(dst_load, src_load);
1057 if (src_load * env->imbalance_pct < dst_load * 100)
1061 task_numa_assign(env, cur, imp);
1066 static void task_numa_find_cpu(struct task_numa_env *env, long imp)
1070 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1071 /* Skip this CPU if the source task cannot migrate */
1072 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1076 task_numa_compare(env, imp);
1080 static int task_numa_migrate(struct task_struct *p)
1082 struct task_numa_env env = {
1085 .src_cpu = task_cpu(p),
1086 .src_nid = cpu_to_node(task_cpu(p)),
1088 .imbalance_pct = 112,
1094 struct sched_domain *sd;
1095 unsigned long faults;
1100 * Pick the lowest SD_NUMA domain, as that would have the smallest
1101 * imbalance and would be the first to start moving tasks about.
1103 * And we want to avoid any moving of tasks about, as that would create
1104 * random movement of tasks -- counter the numa conditions we're trying
1108 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1109 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1112 faults = task_faults(p, env.src_nid);
1113 update_numa_stats(&env.src_stats, env.src_nid);
1114 env.dst_nid = p->numa_preferred_nid;
1115 imp = task_faults(env.p, env.dst_nid) - faults;
1116 update_numa_stats(&env.dst_stats, env.dst_nid);
1118 /* If the preferred nid has capacity, try to use it. */
1119 if (env.dst_stats.has_capacity)
1120 task_numa_find_cpu(&env, imp);
1122 /* No space available on the preferred nid. Look elsewhere. */
1123 if (env.best_cpu == -1) {
1124 for_each_online_node(nid) {
1125 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1128 /* Only consider nodes that recorded more faults */
1129 imp = task_faults(env.p, nid) - faults;
1134 update_numa_stats(&env.dst_stats, env.dst_nid);
1135 task_numa_find_cpu(&env, imp);
1139 /* No better CPU than the current one was found. */
1140 if (env.best_cpu == -1)
1143 if (env.best_task == NULL) {
1144 int ret = migrate_task_to(p, env.best_cpu);
1148 ret = migrate_swap(p, env.best_task);
1149 put_task_struct(env.best_task);
1153 /* Attempt to migrate a task to a CPU on the preferred node. */
1154 static void numa_migrate_preferred(struct task_struct *p)
1156 /* Success if task is already running on preferred CPU */
1157 p->numa_migrate_retry = 0;
1158 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1160 * If migration is temporarily disabled due to a task migration
1161 * then re-enable it now as the task is running on its
1162 * preferred node and memory should migrate locally
1164 if (!p->numa_migrate_seq)
1165 p->numa_migrate_seq++;
1169 /* This task has no NUMA fault statistics yet */
1170 if (unlikely(p->numa_preferred_nid == -1))
1173 /* Otherwise, try migrate to a CPU on the preferred node */
1174 if (task_numa_migrate(p) != 0)
1175 p->numa_migrate_retry = jiffies + HZ*5;
1178 static void task_numa_placement(struct task_struct *p)
1180 int seq, nid, max_nid = -1;
1181 unsigned long max_faults = 0;
1183 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1184 if (p->numa_scan_seq == seq)
1186 p->numa_scan_seq = seq;
1187 p->numa_migrate_seq++;
1188 p->numa_scan_period_max = task_scan_max(p);
1190 /* Find the node with the highest number of faults */
1191 for_each_online_node(nid) {
1192 unsigned long faults = 0;
1195 for (priv = 0; priv < 2; priv++) {
1198 i = task_faults_idx(nid, priv);
1199 diff = -p->numa_faults[i];
1201 /* Decay existing window, copy faults since last scan */
1202 p->numa_faults[i] >>= 1;
1203 p->numa_faults[i] += p->numa_faults_buffer[i];
1204 p->numa_faults_buffer[i] = 0;
1206 faults += p->numa_faults[i];
1207 diff += p->numa_faults[i];
1208 if (p->numa_group) {
1209 /* safe because we can only change our own group */
1210 atomic_long_add(diff, &p->numa_group->faults[i]);
1214 if (faults > max_faults) {
1215 max_faults = faults;
1220 /* Preferred node as the node with the most faults */
1221 if (max_faults && max_nid != p->numa_preferred_nid) {
1222 /* Update the preferred nid and migrate task if possible */
1223 p->numa_preferred_nid = max_nid;
1224 p->numa_migrate_seq = 1;
1225 numa_migrate_preferred(p);
1229 static inline int get_numa_group(struct numa_group *grp)
1231 return atomic_inc_not_zero(&grp->refcount);
1234 static inline void put_numa_group(struct numa_group *grp)
1236 if (atomic_dec_and_test(&grp->refcount))
1237 kfree_rcu(grp, rcu);
1240 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1246 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1249 static void task_numa_group(struct task_struct *p, int cpupid)
1251 struct numa_group *grp, *my_grp;
1252 struct task_struct *tsk;
1254 int cpu = cpupid_to_cpu(cpupid);
1257 if (unlikely(!p->numa_group)) {
1258 unsigned int size = sizeof(struct numa_group) +
1259 2*nr_node_ids*sizeof(atomic_long_t);
1261 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1265 atomic_set(&grp->refcount, 1);
1266 spin_lock_init(&grp->lock);
1267 INIT_LIST_HEAD(&grp->task_list);
1269 for (i = 0; i < 2*nr_node_ids; i++)
1270 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1272 list_add(&p->numa_entry, &grp->task_list);
1274 rcu_assign_pointer(p->numa_group, grp);
1278 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1280 if (!cpupid_match_pid(tsk, cpupid))
1283 grp = rcu_dereference(tsk->numa_group);
1287 my_grp = p->numa_group;
1292 * Only join the other group if its bigger; if we're the bigger group,
1293 * the other task will join us.
1295 if (my_grp->nr_tasks > grp->nr_tasks)
1299 * Tie-break on the grp address.
1301 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1304 if (!get_numa_group(grp))
1315 for (i = 0; i < 2*nr_node_ids; i++) {
1316 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1317 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1320 double_lock(&my_grp->lock, &grp->lock);
1322 list_move(&p->numa_entry, &grp->task_list);
1326 spin_unlock(&my_grp->lock);
1327 spin_unlock(&grp->lock);
1329 rcu_assign_pointer(p->numa_group, grp);
1331 put_numa_group(my_grp);
1334 void task_numa_free(struct task_struct *p)
1336 struct numa_group *grp = p->numa_group;
1340 for (i = 0; i < 2*nr_node_ids; i++)
1341 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1343 spin_lock(&grp->lock);
1344 list_del(&p->numa_entry);
1346 spin_unlock(&grp->lock);
1347 rcu_assign_pointer(p->numa_group, NULL);
1348 put_numa_group(grp);
1351 kfree(p->numa_faults);
1355 * Got a PROT_NONE fault for a page on @node.
1357 void task_numa_fault(int last_cpupid, int node, int pages, bool migrated)
1359 struct task_struct *p = current;
1362 if (!numabalancing_enabled)
1365 /* for example, ksmd faulting in a user's mm */
1369 /* Allocate buffer to track faults on a per-node basis */
1370 if (unlikely(!p->numa_faults)) {
1371 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1373 /* numa_faults and numa_faults_buffer share the allocation */
1374 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1375 if (!p->numa_faults)
1378 BUG_ON(p->numa_faults_buffer);
1379 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1383 * First accesses are treated as private, otherwise consider accesses
1384 * to be private if the accessing pid has not changed
1386 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1389 priv = cpupid_match_pid(p, last_cpupid);
1391 task_numa_group(p, last_cpupid);
1395 * If pages are properly placed (did not migrate) then scan slower.
1396 * This is reset periodically in case of phase changes
1399 /* Initialise if necessary */
1400 if (!p->numa_scan_period_max)
1401 p->numa_scan_period_max = task_scan_max(p);
1403 p->numa_scan_period = min(p->numa_scan_period_max,
1404 p->numa_scan_period + 10);
1407 task_numa_placement(p);
1409 /* Retry task to preferred node migration if it previously failed */
1410 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1411 numa_migrate_preferred(p);
1413 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1416 static void reset_ptenuma_scan(struct task_struct *p)
1418 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1419 p->mm->numa_scan_offset = 0;
1423 * The expensive part of numa migration is done from task_work context.
1424 * Triggered from task_tick_numa().
1426 void task_numa_work(struct callback_head *work)
1428 unsigned long migrate, next_scan, now = jiffies;
1429 struct task_struct *p = current;
1430 struct mm_struct *mm = p->mm;
1431 struct vm_area_struct *vma;
1432 unsigned long start, end;
1433 unsigned long nr_pte_updates = 0;
1436 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1438 work->next = work; /* protect against double add */
1440 * Who cares about NUMA placement when they're dying.
1442 * NOTE: make sure not to dereference p->mm before this check,
1443 * exit_task_work() happens _after_ exit_mm() so we could be called
1444 * without p->mm even though we still had it when we enqueued this
1447 if (p->flags & PF_EXITING)
1450 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1451 mm->numa_next_scan = now +
1452 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1453 mm->numa_next_reset = now +
1454 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1458 * Reset the scan period if enough time has gone by. Objective is that
1459 * scanning will be reduced if pages are properly placed. As tasks
1460 * can enter different phases this needs to be re-examined. Lacking
1461 * proper tracking of reference behaviour, this blunt hammer is used.
1463 migrate = mm->numa_next_reset;
1464 if (time_after(now, migrate)) {
1465 p->numa_scan_period = task_scan_min(p);
1466 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1467 xchg(&mm->numa_next_reset, next_scan);
1471 * Enforce maximal scan/migration frequency..
1473 migrate = mm->numa_next_scan;
1474 if (time_before(now, migrate))
1477 if (p->numa_scan_period == 0) {
1478 p->numa_scan_period_max = task_scan_max(p);
1479 p->numa_scan_period = task_scan_min(p);
1482 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1483 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1487 * Delay this task enough that another task of this mm will likely win
1488 * the next time around.
1490 p->node_stamp += 2 * TICK_NSEC;
1492 start = mm->numa_scan_offset;
1493 pages = sysctl_numa_balancing_scan_size;
1494 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1498 down_read(&mm->mmap_sem);
1499 vma = find_vma(mm, start);
1501 reset_ptenuma_scan(p);
1505 for (; vma; vma = vma->vm_next) {
1506 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1510 * Shared library pages mapped by multiple processes are not
1511 * migrated as it is expected they are cache replicated. Avoid
1512 * hinting faults in read-only file-backed mappings or the vdso
1513 * as migrating the pages will be of marginal benefit.
1516 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1520 start = max(start, vma->vm_start);
1521 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1522 end = min(end, vma->vm_end);
1523 nr_pte_updates += change_prot_numa(vma, start, end);
1526 * Scan sysctl_numa_balancing_scan_size but ensure that
1527 * at least one PTE is updated so that unused virtual
1528 * address space is quickly skipped.
1531 pages -= (end - start) >> PAGE_SHIFT;
1536 } while (end != vma->vm_end);
1541 * If the whole process was scanned without updates then no NUMA
1542 * hinting faults are being recorded and scan rate should be lower.
1544 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1545 p->numa_scan_period = min(p->numa_scan_period_max,
1546 p->numa_scan_period << 1);
1548 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1549 mm->numa_next_scan = next_scan;
1553 * It is possible to reach the end of the VMA list but the last few
1554 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1555 * would find the !migratable VMA on the next scan but not reset the
1556 * scanner to the start so check it now.
1559 mm->numa_scan_offset = start;
1561 reset_ptenuma_scan(p);
1562 up_read(&mm->mmap_sem);
1566 * Drive the periodic memory faults..
1568 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1570 struct callback_head *work = &curr->numa_work;
1574 * We don't care about NUMA placement if we don't have memory.
1576 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1580 * Using runtime rather than walltime has the dual advantage that
1581 * we (mostly) drive the selection from busy threads and that the
1582 * task needs to have done some actual work before we bother with
1585 now = curr->se.sum_exec_runtime;
1586 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1588 if (now - curr->node_stamp > period) {
1589 if (!curr->node_stamp)
1590 curr->numa_scan_period = task_scan_min(curr);
1591 curr->node_stamp += period;
1593 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1594 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1595 task_work_add(curr, work, true);
1600 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1603 #endif /* CONFIG_NUMA_BALANCING */
1606 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1608 update_load_add(&cfs_rq->load, se->load.weight);
1609 if (!parent_entity(se))
1610 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1612 if (entity_is_task(se))
1613 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1615 cfs_rq->nr_running++;
1619 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1621 update_load_sub(&cfs_rq->load, se->load.weight);
1622 if (!parent_entity(se))
1623 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1624 if (entity_is_task(se))
1625 list_del_init(&se->group_node);
1626 cfs_rq->nr_running--;
1629 #ifdef CONFIG_FAIR_GROUP_SCHED
1631 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1636 * Use this CPU's actual weight instead of the last load_contribution
1637 * to gain a more accurate current total weight. See
1638 * update_cfs_rq_load_contribution().
1640 tg_weight = atomic_long_read(&tg->load_avg);
1641 tg_weight -= cfs_rq->tg_load_contrib;
1642 tg_weight += cfs_rq->load.weight;
1647 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1649 long tg_weight, load, shares;
1651 tg_weight = calc_tg_weight(tg, cfs_rq);
1652 load = cfs_rq->load.weight;
1654 shares = (tg->shares * load);
1656 shares /= tg_weight;
1658 if (shares < MIN_SHARES)
1659 shares = MIN_SHARES;
1660 if (shares > tg->shares)
1661 shares = tg->shares;
1665 # else /* CONFIG_SMP */
1666 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1670 # endif /* CONFIG_SMP */
1671 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1672 unsigned long weight)
1675 /* commit outstanding execution time */
1676 if (cfs_rq->curr == se)
1677 update_curr(cfs_rq);
1678 account_entity_dequeue(cfs_rq, se);
1681 update_load_set(&se->load, weight);
1684 account_entity_enqueue(cfs_rq, se);
1687 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1689 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1691 struct task_group *tg;
1692 struct sched_entity *se;
1696 se = tg->se[cpu_of(rq_of(cfs_rq))];
1697 if (!se || throttled_hierarchy(cfs_rq))
1700 if (likely(se->load.weight == tg->shares))
1703 shares = calc_cfs_shares(cfs_rq, tg);
1705 reweight_entity(cfs_rq_of(se), se, shares);
1707 #else /* CONFIG_FAIR_GROUP_SCHED */
1708 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1711 #endif /* CONFIG_FAIR_GROUP_SCHED */
1715 * We choose a half-life close to 1 scheduling period.
1716 * Note: The tables below are dependent on this value.
1718 #define LOAD_AVG_PERIOD 32
1719 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1720 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1722 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1723 static const u32 runnable_avg_yN_inv[] = {
1724 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1725 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1726 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1727 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1728 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1729 0x85aac367, 0x82cd8698,
1733 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1734 * over-estimates when re-combining.
1736 static const u32 runnable_avg_yN_sum[] = {
1737 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1738 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1739 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1744 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1746 static __always_inline u64 decay_load(u64 val, u64 n)
1748 unsigned int local_n;
1752 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1755 /* after bounds checking we can collapse to 32-bit */
1759 * As y^PERIOD = 1/2, we can combine
1760 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1761 * With a look-up table which covers k^n (n<PERIOD)
1763 * To achieve constant time decay_load.
1765 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1766 val >>= local_n / LOAD_AVG_PERIOD;
1767 local_n %= LOAD_AVG_PERIOD;
1770 val *= runnable_avg_yN_inv[local_n];
1771 /* We don't use SRR here since we always want to round down. */
1776 * For updates fully spanning n periods, the contribution to runnable
1777 * average will be: \Sum 1024*y^n
1779 * We can compute this reasonably efficiently by combining:
1780 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1782 static u32 __compute_runnable_contrib(u64 n)
1786 if (likely(n <= LOAD_AVG_PERIOD))
1787 return runnable_avg_yN_sum[n];
1788 else if (unlikely(n >= LOAD_AVG_MAX_N))
1789 return LOAD_AVG_MAX;
1791 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1793 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1794 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1796 n -= LOAD_AVG_PERIOD;
1797 } while (n > LOAD_AVG_PERIOD);
1799 contrib = decay_load(contrib, n);
1800 return contrib + runnable_avg_yN_sum[n];
1804 * We can represent the historical contribution to runnable average as the
1805 * coefficients of a geometric series. To do this we sub-divide our runnable
1806 * history into segments of approximately 1ms (1024us); label the segment that
1807 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1809 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1811 * (now) (~1ms ago) (~2ms ago)
1813 * Let u_i denote the fraction of p_i that the entity was runnable.
1815 * We then designate the fractions u_i as our co-efficients, yielding the
1816 * following representation of historical load:
1817 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1819 * We choose y based on the with of a reasonably scheduling period, fixing:
1822 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1823 * approximately half as much as the contribution to load within the last ms
1826 * When a period "rolls over" and we have new u_0`, multiplying the previous
1827 * sum again by y is sufficient to update:
1828 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1829 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1831 static __always_inline int __update_entity_runnable_avg(u64 now,
1832 struct sched_avg *sa,
1836 u32 runnable_contrib;
1837 int delta_w, decayed = 0;
1839 delta = now - sa->last_runnable_update;
1841 * This should only happen when time goes backwards, which it
1842 * unfortunately does during sched clock init when we swap over to TSC.
1844 if ((s64)delta < 0) {
1845 sa->last_runnable_update = now;
1850 * Use 1024ns as the unit of measurement since it's a reasonable
1851 * approximation of 1us and fast to compute.
1856 sa->last_runnable_update = now;
1858 /* delta_w is the amount already accumulated against our next period */
1859 delta_w = sa->runnable_avg_period % 1024;
1860 if (delta + delta_w >= 1024) {
1861 /* period roll-over */
1865 * Now that we know we're crossing a period boundary, figure
1866 * out how much from delta we need to complete the current
1867 * period and accrue it.
1869 delta_w = 1024 - delta_w;
1871 sa->runnable_avg_sum += delta_w;
1872 sa->runnable_avg_period += delta_w;
1876 /* Figure out how many additional periods this update spans */
1877 periods = delta / 1024;
1880 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1882 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1885 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1886 runnable_contrib = __compute_runnable_contrib(periods);
1888 sa->runnable_avg_sum += runnable_contrib;
1889 sa->runnable_avg_period += runnable_contrib;
1892 /* Remainder of delta accrued against u_0` */
1894 sa->runnable_avg_sum += delta;
1895 sa->runnable_avg_period += delta;
1900 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1901 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1904 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1906 decays -= se->avg.decay_count;
1910 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1911 se->avg.decay_count = 0;
1916 #ifdef CONFIG_FAIR_GROUP_SCHED
1917 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1920 struct task_group *tg = cfs_rq->tg;
1923 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1924 tg_contrib -= cfs_rq->tg_load_contrib;
1926 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1927 atomic_long_add(tg_contrib, &tg->load_avg);
1928 cfs_rq->tg_load_contrib += tg_contrib;
1933 * Aggregate cfs_rq runnable averages into an equivalent task_group
1934 * representation for computing load contributions.
1936 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1937 struct cfs_rq *cfs_rq)
1939 struct task_group *tg = cfs_rq->tg;
1942 /* The fraction of a cpu used by this cfs_rq */
1943 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1944 sa->runnable_avg_period + 1);
1945 contrib -= cfs_rq->tg_runnable_contrib;
1947 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1948 atomic_add(contrib, &tg->runnable_avg);
1949 cfs_rq->tg_runnable_contrib += contrib;
1953 static inline void __update_group_entity_contrib(struct sched_entity *se)
1955 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1956 struct task_group *tg = cfs_rq->tg;
1961 contrib = cfs_rq->tg_load_contrib * tg->shares;
1962 se->avg.load_avg_contrib = div_u64(contrib,
1963 atomic_long_read(&tg->load_avg) + 1);
1966 * For group entities we need to compute a correction term in the case
1967 * that they are consuming <1 cpu so that we would contribute the same
1968 * load as a task of equal weight.
1970 * Explicitly co-ordinating this measurement would be expensive, but
1971 * fortunately the sum of each cpus contribution forms a usable
1972 * lower-bound on the true value.
1974 * Consider the aggregate of 2 contributions. Either they are disjoint
1975 * (and the sum represents true value) or they are disjoint and we are
1976 * understating by the aggregate of their overlap.
1978 * Extending this to N cpus, for a given overlap, the maximum amount we
1979 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1980 * cpus that overlap for this interval and w_i is the interval width.
1982 * On a small machine; the first term is well-bounded which bounds the
1983 * total error since w_i is a subset of the period. Whereas on a
1984 * larger machine, while this first term can be larger, if w_i is the
1985 * of consequential size guaranteed to see n_i*w_i quickly converge to
1986 * our upper bound of 1-cpu.
1988 runnable_avg = atomic_read(&tg->runnable_avg);
1989 if (runnable_avg < NICE_0_LOAD) {
1990 se->avg.load_avg_contrib *= runnable_avg;
1991 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1995 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1996 int force_update) {}
1997 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1998 struct cfs_rq *cfs_rq) {}
1999 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2002 static inline void __update_task_entity_contrib(struct sched_entity *se)
2006 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2007 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2008 contrib /= (se->avg.runnable_avg_period + 1);
2009 se->avg.load_avg_contrib = scale_load(contrib);
2012 /* Compute the current contribution to load_avg by se, return any delta */
2013 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2015 long old_contrib = se->avg.load_avg_contrib;
2017 if (entity_is_task(se)) {
2018 __update_task_entity_contrib(se);
2020 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2021 __update_group_entity_contrib(se);
2024 return se->avg.load_avg_contrib - old_contrib;
2027 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2030 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2031 cfs_rq->blocked_load_avg -= load_contrib;
2033 cfs_rq->blocked_load_avg = 0;
2036 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2038 /* Update a sched_entity's runnable average */
2039 static inline void update_entity_load_avg(struct sched_entity *se,
2042 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2047 * For a group entity we need to use their owned cfs_rq_clock_task() in
2048 * case they are the parent of a throttled hierarchy.
2050 if (entity_is_task(se))
2051 now = cfs_rq_clock_task(cfs_rq);
2053 now = cfs_rq_clock_task(group_cfs_rq(se));
2055 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2058 contrib_delta = __update_entity_load_avg_contrib(se);
2064 cfs_rq->runnable_load_avg += contrib_delta;
2066 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2070 * Decay the load contributed by all blocked children and account this so that
2071 * their contribution may appropriately discounted when they wake up.
2073 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2075 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2078 decays = now - cfs_rq->last_decay;
2079 if (!decays && !force_update)
2082 if (atomic_long_read(&cfs_rq->removed_load)) {
2083 unsigned long removed_load;
2084 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2085 subtract_blocked_load_contrib(cfs_rq, removed_load);
2089 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2091 atomic64_add(decays, &cfs_rq->decay_counter);
2092 cfs_rq->last_decay = now;
2095 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2098 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2100 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2101 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2104 /* Add the load generated by se into cfs_rq's child load-average */
2105 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2106 struct sched_entity *se,
2110 * We track migrations using entity decay_count <= 0, on a wake-up
2111 * migration we use a negative decay count to track the remote decays
2112 * accumulated while sleeping.
2114 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2115 * are seen by enqueue_entity_load_avg() as a migration with an already
2116 * constructed load_avg_contrib.
2118 if (unlikely(se->avg.decay_count <= 0)) {
2119 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2120 if (se->avg.decay_count) {
2122 * In a wake-up migration we have to approximate the
2123 * time sleeping. This is because we can't synchronize
2124 * clock_task between the two cpus, and it is not
2125 * guaranteed to be read-safe. Instead, we can
2126 * approximate this using our carried decays, which are
2127 * explicitly atomically readable.
2129 se->avg.last_runnable_update -= (-se->avg.decay_count)
2131 update_entity_load_avg(se, 0);
2132 /* Indicate that we're now synchronized and on-rq */
2133 se->avg.decay_count = 0;
2138 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2139 * would have made count negative); we must be careful to avoid
2140 * double-accounting blocked time after synchronizing decays.
2142 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2146 /* migrated tasks did not contribute to our blocked load */
2148 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2149 update_entity_load_avg(se, 0);
2152 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2153 /* we force update consideration on load-balancer moves */
2154 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2158 * Remove se's load from this cfs_rq child load-average, if the entity is
2159 * transitioning to a blocked state we track its projected decay using
2162 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2163 struct sched_entity *se,
2166 update_entity_load_avg(se, 1);
2167 /* we force update consideration on load-balancer moves */
2168 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2170 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2172 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2173 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2174 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2178 * Update the rq's load with the elapsed running time before entering
2179 * idle. if the last scheduled task is not a CFS task, idle_enter will
2180 * be the only way to update the runnable statistic.
2182 void idle_enter_fair(struct rq *this_rq)
2184 update_rq_runnable_avg(this_rq, 1);
2188 * Update the rq's load with the elapsed idle time before a task is
2189 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2190 * be the only way to update the runnable statistic.
2192 void idle_exit_fair(struct rq *this_rq)
2194 update_rq_runnable_avg(this_rq, 0);
2198 static inline void update_entity_load_avg(struct sched_entity *se,
2199 int update_cfs_rq) {}
2200 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2201 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2202 struct sched_entity *se,
2204 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2205 struct sched_entity *se,
2207 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2208 int force_update) {}
2211 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2213 #ifdef CONFIG_SCHEDSTATS
2214 struct task_struct *tsk = NULL;
2216 if (entity_is_task(se))
2219 if (se->statistics.sleep_start) {
2220 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2225 if (unlikely(delta > se->statistics.sleep_max))
2226 se->statistics.sleep_max = delta;
2228 se->statistics.sleep_start = 0;
2229 se->statistics.sum_sleep_runtime += delta;
2232 account_scheduler_latency(tsk, delta >> 10, 1);
2233 trace_sched_stat_sleep(tsk, delta);
2236 if (se->statistics.block_start) {
2237 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2242 if (unlikely(delta > se->statistics.block_max))
2243 se->statistics.block_max = delta;
2245 se->statistics.block_start = 0;
2246 se->statistics.sum_sleep_runtime += delta;
2249 if (tsk->in_iowait) {
2250 se->statistics.iowait_sum += delta;
2251 se->statistics.iowait_count++;
2252 trace_sched_stat_iowait(tsk, delta);
2255 trace_sched_stat_blocked(tsk, delta);
2258 * Blocking time is in units of nanosecs, so shift by
2259 * 20 to get a milliseconds-range estimation of the
2260 * amount of time that the task spent sleeping:
2262 if (unlikely(prof_on == SLEEP_PROFILING)) {
2263 profile_hits(SLEEP_PROFILING,
2264 (void *)get_wchan(tsk),
2267 account_scheduler_latency(tsk, delta >> 10, 0);
2273 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2275 #ifdef CONFIG_SCHED_DEBUG
2276 s64 d = se->vruntime - cfs_rq->min_vruntime;
2281 if (d > 3*sysctl_sched_latency)
2282 schedstat_inc(cfs_rq, nr_spread_over);
2287 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2289 u64 vruntime = cfs_rq->min_vruntime;
2292 * The 'current' period is already promised to the current tasks,
2293 * however the extra weight of the new task will slow them down a
2294 * little, place the new task so that it fits in the slot that
2295 * stays open at the end.
2297 if (initial && sched_feat(START_DEBIT))
2298 vruntime += sched_vslice(cfs_rq, se);
2300 /* sleeps up to a single latency don't count. */
2302 unsigned long thresh = sysctl_sched_latency;
2305 * Halve their sleep time's effect, to allow
2306 * for a gentler effect of sleepers:
2308 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2314 /* ensure we never gain time by being placed backwards. */
2315 se->vruntime = max_vruntime(se->vruntime, vruntime);
2318 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2321 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2324 * Update the normalized vruntime before updating min_vruntime
2325 * through calling update_curr().
2327 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2328 se->vruntime += cfs_rq->min_vruntime;
2331 * Update run-time statistics of the 'current'.
2333 update_curr(cfs_rq);
2334 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2335 account_entity_enqueue(cfs_rq, se);
2336 update_cfs_shares(cfs_rq);
2338 if (flags & ENQUEUE_WAKEUP) {
2339 place_entity(cfs_rq, se, 0);
2340 enqueue_sleeper(cfs_rq, se);
2343 update_stats_enqueue(cfs_rq, se);
2344 check_spread(cfs_rq, se);
2345 if (se != cfs_rq->curr)
2346 __enqueue_entity(cfs_rq, se);
2349 if (cfs_rq->nr_running == 1) {
2350 list_add_leaf_cfs_rq(cfs_rq);
2351 check_enqueue_throttle(cfs_rq);
2355 static void __clear_buddies_last(struct sched_entity *se)
2357 for_each_sched_entity(se) {
2358 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2359 if (cfs_rq->last == se)
2360 cfs_rq->last = NULL;
2366 static void __clear_buddies_next(struct sched_entity *se)
2368 for_each_sched_entity(se) {
2369 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2370 if (cfs_rq->next == se)
2371 cfs_rq->next = NULL;
2377 static void __clear_buddies_skip(struct sched_entity *se)
2379 for_each_sched_entity(se) {
2380 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2381 if (cfs_rq->skip == se)
2382 cfs_rq->skip = NULL;
2388 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2390 if (cfs_rq->last == se)
2391 __clear_buddies_last(se);
2393 if (cfs_rq->next == se)
2394 __clear_buddies_next(se);
2396 if (cfs_rq->skip == se)
2397 __clear_buddies_skip(se);
2400 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2403 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2406 * Update run-time statistics of the 'current'.
2408 update_curr(cfs_rq);
2409 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2411 update_stats_dequeue(cfs_rq, se);
2412 if (flags & DEQUEUE_SLEEP) {
2413 #ifdef CONFIG_SCHEDSTATS
2414 if (entity_is_task(se)) {
2415 struct task_struct *tsk = task_of(se);
2417 if (tsk->state & TASK_INTERRUPTIBLE)
2418 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2419 if (tsk->state & TASK_UNINTERRUPTIBLE)
2420 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2425 clear_buddies(cfs_rq, se);
2427 if (se != cfs_rq->curr)
2428 __dequeue_entity(cfs_rq, se);
2430 account_entity_dequeue(cfs_rq, se);
2433 * Normalize the entity after updating the min_vruntime because the
2434 * update can refer to the ->curr item and we need to reflect this
2435 * movement in our normalized position.
2437 if (!(flags & DEQUEUE_SLEEP))
2438 se->vruntime -= cfs_rq->min_vruntime;
2440 /* return excess runtime on last dequeue */
2441 return_cfs_rq_runtime(cfs_rq);
2443 update_min_vruntime(cfs_rq);
2444 update_cfs_shares(cfs_rq);
2448 * Preempt the current task with a newly woken task if needed:
2451 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2453 unsigned long ideal_runtime, delta_exec;
2454 struct sched_entity *se;
2457 ideal_runtime = sched_slice(cfs_rq, curr);
2458 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2459 if (delta_exec > ideal_runtime) {
2460 resched_task(rq_of(cfs_rq)->curr);
2462 * The current task ran long enough, ensure it doesn't get
2463 * re-elected due to buddy favours.
2465 clear_buddies(cfs_rq, curr);
2470 * Ensure that a task that missed wakeup preemption by a
2471 * narrow margin doesn't have to wait for a full slice.
2472 * This also mitigates buddy induced latencies under load.
2474 if (delta_exec < sysctl_sched_min_granularity)
2477 se = __pick_first_entity(cfs_rq);
2478 delta = curr->vruntime - se->vruntime;
2483 if (delta > ideal_runtime)
2484 resched_task(rq_of(cfs_rq)->curr);
2488 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2490 /* 'current' is not kept within the tree. */
2493 * Any task has to be enqueued before it get to execute on
2494 * a CPU. So account for the time it spent waiting on the
2497 update_stats_wait_end(cfs_rq, se);
2498 __dequeue_entity(cfs_rq, se);
2501 update_stats_curr_start(cfs_rq, se);
2503 #ifdef CONFIG_SCHEDSTATS
2505 * Track our maximum slice length, if the CPU's load is at
2506 * least twice that of our own weight (i.e. dont track it
2507 * when there are only lesser-weight tasks around):
2509 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2510 se->statistics.slice_max = max(se->statistics.slice_max,
2511 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2514 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2518 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2521 * Pick the next process, keeping these things in mind, in this order:
2522 * 1) keep things fair between processes/task groups
2523 * 2) pick the "next" process, since someone really wants that to run
2524 * 3) pick the "last" process, for cache locality
2525 * 4) do not run the "skip" process, if something else is available
2527 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2529 struct sched_entity *se = __pick_first_entity(cfs_rq);
2530 struct sched_entity *left = se;
2533 * Avoid running the skip buddy, if running something else can
2534 * be done without getting too unfair.
2536 if (cfs_rq->skip == se) {
2537 struct sched_entity *second = __pick_next_entity(se);
2538 if (second && wakeup_preempt_entity(second, left) < 1)
2543 * Prefer last buddy, try to return the CPU to a preempted task.
2545 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2549 * Someone really wants this to run. If it's not unfair, run it.
2551 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2554 clear_buddies(cfs_rq, se);
2559 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2561 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2564 * If still on the runqueue then deactivate_task()
2565 * was not called and update_curr() has to be done:
2568 update_curr(cfs_rq);
2570 /* throttle cfs_rqs exceeding runtime */
2571 check_cfs_rq_runtime(cfs_rq);
2573 check_spread(cfs_rq, prev);
2575 update_stats_wait_start(cfs_rq, prev);
2576 /* Put 'current' back into the tree. */
2577 __enqueue_entity(cfs_rq, prev);
2578 /* in !on_rq case, update occurred at dequeue */
2579 update_entity_load_avg(prev, 1);
2581 cfs_rq->curr = NULL;
2585 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2588 * Update run-time statistics of the 'current'.
2590 update_curr(cfs_rq);
2593 * Ensure that runnable average is periodically updated.
2595 update_entity_load_avg(curr, 1);
2596 update_cfs_rq_blocked_load(cfs_rq, 1);
2597 update_cfs_shares(cfs_rq);
2599 #ifdef CONFIG_SCHED_HRTICK
2601 * queued ticks are scheduled to match the slice, so don't bother
2602 * validating it and just reschedule.
2605 resched_task(rq_of(cfs_rq)->curr);
2609 * don't let the period tick interfere with the hrtick preemption
2611 if (!sched_feat(DOUBLE_TICK) &&
2612 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2616 if (cfs_rq->nr_running > 1)
2617 check_preempt_tick(cfs_rq, curr);
2621 /**************************************************
2622 * CFS bandwidth control machinery
2625 #ifdef CONFIG_CFS_BANDWIDTH
2627 #ifdef HAVE_JUMP_LABEL
2628 static struct static_key __cfs_bandwidth_used;
2630 static inline bool cfs_bandwidth_used(void)
2632 return static_key_false(&__cfs_bandwidth_used);
2635 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2637 /* only need to count groups transitioning between enabled/!enabled */
2638 if (enabled && !was_enabled)
2639 static_key_slow_inc(&__cfs_bandwidth_used);
2640 else if (!enabled && was_enabled)
2641 static_key_slow_dec(&__cfs_bandwidth_used);
2643 #else /* HAVE_JUMP_LABEL */
2644 static bool cfs_bandwidth_used(void)
2649 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2650 #endif /* HAVE_JUMP_LABEL */
2653 * default period for cfs group bandwidth.
2654 * default: 0.1s, units: nanoseconds
2656 static inline u64 default_cfs_period(void)
2658 return 100000000ULL;
2661 static inline u64 sched_cfs_bandwidth_slice(void)
2663 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2667 * Replenish runtime according to assigned quota and update expiration time.
2668 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2669 * additional synchronization around rq->lock.
2671 * requires cfs_b->lock
2673 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2677 if (cfs_b->quota == RUNTIME_INF)
2680 now = sched_clock_cpu(smp_processor_id());
2681 cfs_b->runtime = cfs_b->quota;
2682 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2685 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2687 return &tg->cfs_bandwidth;
2690 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2691 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2693 if (unlikely(cfs_rq->throttle_count))
2694 return cfs_rq->throttled_clock_task;
2696 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2699 /* returns 0 on failure to allocate runtime */
2700 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2702 struct task_group *tg = cfs_rq->tg;
2703 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2704 u64 amount = 0, min_amount, expires;
2706 /* note: this is a positive sum as runtime_remaining <= 0 */
2707 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2709 raw_spin_lock(&cfs_b->lock);
2710 if (cfs_b->quota == RUNTIME_INF)
2711 amount = min_amount;
2714 * If the bandwidth pool has become inactive, then at least one
2715 * period must have elapsed since the last consumption.
2716 * Refresh the global state and ensure bandwidth timer becomes
2719 if (!cfs_b->timer_active) {
2720 __refill_cfs_bandwidth_runtime(cfs_b);
2721 __start_cfs_bandwidth(cfs_b);
2724 if (cfs_b->runtime > 0) {
2725 amount = min(cfs_b->runtime, min_amount);
2726 cfs_b->runtime -= amount;
2730 expires = cfs_b->runtime_expires;
2731 raw_spin_unlock(&cfs_b->lock);
2733 cfs_rq->runtime_remaining += amount;
2735 * we may have advanced our local expiration to account for allowed
2736 * spread between our sched_clock and the one on which runtime was
2739 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2740 cfs_rq->runtime_expires = expires;
2742 return cfs_rq->runtime_remaining > 0;
2746 * Note: This depends on the synchronization provided by sched_clock and the
2747 * fact that rq->clock snapshots this value.
2749 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2751 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2753 /* if the deadline is ahead of our clock, nothing to do */
2754 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2757 if (cfs_rq->runtime_remaining < 0)
2761 * If the local deadline has passed we have to consider the
2762 * possibility that our sched_clock is 'fast' and the global deadline
2763 * has not truly expired.
2765 * Fortunately we can check determine whether this the case by checking
2766 * whether the global deadline has advanced.
2769 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2770 /* extend local deadline, drift is bounded above by 2 ticks */
2771 cfs_rq->runtime_expires += TICK_NSEC;
2773 /* global deadline is ahead, expiration has passed */
2774 cfs_rq->runtime_remaining = 0;
2778 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2779 unsigned long delta_exec)
2781 /* dock delta_exec before expiring quota (as it could span periods) */
2782 cfs_rq->runtime_remaining -= delta_exec;
2783 expire_cfs_rq_runtime(cfs_rq);
2785 if (likely(cfs_rq->runtime_remaining > 0))
2789 * if we're unable to extend our runtime we resched so that the active
2790 * hierarchy can be throttled
2792 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2793 resched_task(rq_of(cfs_rq)->curr);
2796 static __always_inline
2797 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2799 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2802 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2805 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2807 return cfs_bandwidth_used() && cfs_rq->throttled;
2810 /* check whether cfs_rq, or any parent, is throttled */
2811 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2813 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2817 * Ensure that neither of the group entities corresponding to src_cpu or
2818 * dest_cpu are members of a throttled hierarchy when performing group
2819 * load-balance operations.
2821 static inline int throttled_lb_pair(struct task_group *tg,
2822 int src_cpu, int dest_cpu)
2824 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2826 src_cfs_rq = tg->cfs_rq[src_cpu];
2827 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2829 return throttled_hierarchy(src_cfs_rq) ||
2830 throttled_hierarchy(dest_cfs_rq);
2833 /* updated child weight may affect parent so we have to do this bottom up */
2834 static int tg_unthrottle_up(struct task_group *tg, void *data)
2836 struct rq *rq = data;
2837 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2839 cfs_rq->throttle_count--;
2841 if (!cfs_rq->throttle_count) {
2842 /* adjust cfs_rq_clock_task() */
2843 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2844 cfs_rq->throttled_clock_task;
2851 static int tg_throttle_down(struct task_group *tg, void *data)
2853 struct rq *rq = data;
2854 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2856 /* group is entering throttled state, stop time */
2857 if (!cfs_rq->throttle_count)
2858 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2859 cfs_rq->throttle_count++;
2864 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2866 struct rq *rq = rq_of(cfs_rq);
2867 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2868 struct sched_entity *se;
2869 long task_delta, dequeue = 1;
2871 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2873 /* freeze hierarchy runnable averages while throttled */
2875 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2878 task_delta = cfs_rq->h_nr_running;
2879 for_each_sched_entity(se) {
2880 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2881 /* throttled entity or throttle-on-deactivate */
2886 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2887 qcfs_rq->h_nr_running -= task_delta;
2889 if (qcfs_rq->load.weight)
2894 rq->nr_running -= task_delta;
2896 cfs_rq->throttled = 1;
2897 cfs_rq->throttled_clock = rq_clock(rq);
2898 raw_spin_lock(&cfs_b->lock);
2899 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2900 raw_spin_unlock(&cfs_b->lock);
2903 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2905 struct rq *rq = rq_of(cfs_rq);
2906 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2907 struct sched_entity *se;
2911 se = cfs_rq->tg->se[cpu_of(rq)];
2913 cfs_rq->throttled = 0;
2915 update_rq_clock(rq);
2917 raw_spin_lock(&cfs_b->lock);
2918 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2919 list_del_rcu(&cfs_rq->throttled_list);
2920 raw_spin_unlock(&cfs_b->lock);
2922 /* update hierarchical throttle state */
2923 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2925 if (!cfs_rq->load.weight)
2928 task_delta = cfs_rq->h_nr_running;
2929 for_each_sched_entity(se) {
2933 cfs_rq = cfs_rq_of(se);
2935 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2936 cfs_rq->h_nr_running += task_delta;
2938 if (cfs_rq_throttled(cfs_rq))
2943 rq->nr_running += task_delta;
2945 /* determine whether we need to wake up potentially idle cpu */
2946 if (rq->curr == rq->idle && rq->cfs.nr_running)
2947 resched_task(rq->curr);
2950 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2951 u64 remaining, u64 expires)
2953 struct cfs_rq *cfs_rq;
2954 u64 runtime = remaining;
2957 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2959 struct rq *rq = rq_of(cfs_rq);
2961 raw_spin_lock(&rq->lock);
2962 if (!cfs_rq_throttled(cfs_rq))
2965 runtime = -cfs_rq->runtime_remaining + 1;
2966 if (runtime > remaining)
2967 runtime = remaining;
2968 remaining -= runtime;
2970 cfs_rq->runtime_remaining += runtime;
2971 cfs_rq->runtime_expires = expires;
2973 /* we check whether we're throttled above */
2974 if (cfs_rq->runtime_remaining > 0)
2975 unthrottle_cfs_rq(cfs_rq);
2978 raw_spin_unlock(&rq->lock);
2989 * Responsible for refilling a task_group's bandwidth and unthrottling its
2990 * cfs_rqs as appropriate. If there has been no activity within the last
2991 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2992 * used to track this state.
2994 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2996 u64 runtime, runtime_expires;
2997 int idle = 1, throttled;
2999 raw_spin_lock(&cfs_b->lock);
3000 /* no need to continue the timer with no bandwidth constraint */
3001 if (cfs_b->quota == RUNTIME_INF)
3004 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3005 /* idle depends on !throttled (for the case of a large deficit) */
3006 idle = cfs_b->idle && !throttled;
3007 cfs_b->nr_periods += overrun;
3009 /* if we're going inactive then everything else can be deferred */
3013 __refill_cfs_bandwidth_runtime(cfs_b);
3016 /* mark as potentially idle for the upcoming period */
3021 /* account preceding periods in which throttling occurred */
3022 cfs_b->nr_throttled += overrun;
3025 * There are throttled entities so we must first use the new bandwidth
3026 * to unthrottle them before making it generally available. This
3027 * ensures that all existing debts will be paid before a new cfs_rq is
3030 runtime = cfs_b->runtime;
3031 runtime_expires = cfs_b->runtime_expires;
3035 * This check is repeated as we are holding onto the new bandwidth
3036 * while we unthrottle. This can potentially race with an unthrottled
3037 * group trying to acquire new bandwidth from the global pool.
3039 while (throttled && runtime > 0) {
3040 raw_spin_unlock(&cfs_b->lock);
3041 /* we can't nest cfs_b->lock while distributing bandwidth */
3042 runtime = distribute_cfs_runtime(cfs_b, runtime,
3044 raw_spin_lock(&cfs_b->lock);
3046 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3049 /* return (any) remaining runtime */
3050 cfs_b->runtime = runtime;
3052 * While we are ensured activity in the period following an
3053 * unthrottle, this also covers the case in which the new bandwidth is
3054 * insufficient to cover the existing bandwidth deficit. (Forcing the
3055 * timer to remain active while there are any throttled entities.)
3060 cfs_b->timer_active = 0;
3061 raw_spin_unlock(&cfs_b->lock);
3066 /* a cfs_rq won't donate quota below this amount */
3067 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3068 /* minimum remaining period time to redistribute slack quota */
3069 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3070 /* how long we wait to gather additional slack before distributing */
3071 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3073 /* are we near the end of the current quota period? */
3074 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3076 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3079 /* if the call-back is running a quota refresh is already occurring */
3080 if (hrtimer_callback_running(refresh_timer))
3083 /* is a quota refresh about to occur? */
3084 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3085 if (remaining < min_expire)
3091 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3093 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3095 /* if there's a quota refresh soon don't bother with slack */
3096 if (runtime_refresh_within(cfs_b, min_left))
3099 start_bandwidth_timer(&cfs_b->slack_timer,
3100 ns_to_ktime(cfs_bandwidth_slack_period));
3103 /* we know any runtime found here is valid as update_curr() precedes return */
3104 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3106 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3107 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3109 if (slack_runtime <= 0)
3112 raw_spin_lock(&cfs_b->lock);
3113 if (cfs_b->quota != RUNTIME_INF &&
3114 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3115 cfs_b->runtime += slack_runtime;
3117 /* we are under rq->lock, defer unthrottling using a timer */
3118 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3119 !list_empty(&cfs_b->throttled_cfs_rq))
3120 start_cfs_slack_bandwidth(cfs_b);
3122 raw_spin_unlock(&cfs_b->lock);
3124 /* even if it's not valid for return we don't want to try again */
3125 cfs_rq->runtime_remaining -= slack_runtime;
3128 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3130 if (!cfs_bandwidth_used())
3133 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3136 __return_cfs_rq_runtime(cfs_rq);
3140 * This is done with a timer (instead of inline with bandwidth return) since
3141 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3143 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3145 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3148 /* confirm we're still not at a refresh boundary */
3149 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3152 raw_spin_lock(&cfs_b->lock);
3153 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3154 runtime = cfs_b->runtime;
3157 expires = cfs_b->runtime_expires;
3158 raw_spin_unlock(&cfs_b->lock);
3163 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3165 raw_spin_lock(&cfs_b->lock);
3166 if (expires == cfs_b->runtime_expires)
3167 cfs_b->runtime = runtime;
3168 raw_spin_unlock(&cfs_b->lock);
3172 * When a group wakes up we want to make sure that its quota is not already
3173 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3174 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3176 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3178 if (!cfs_bandwidth_used())
3181 /* an active group must be handled by the update_curr()->put() path */
3182 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3185 /* ensure the group is not already throttled */
3186 if (cfs_rq_throttled(cfs_rq))
3189 /* update runtime allocation */
3190 account_cfs_rq_runtime(cfs_rq, 0);
3191 if (cfs_rq->runtime_remaining <= 0)
3192 throttle_cfs_rq(cfs_rq);
3195 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3196 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3198 if (!cfs_bandwidth_used())
3201 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3205 * it's possible for a throttled entity to be forced into a running
3206 * state (e.g. set_curr_task), in this case we're finished.
3208 if (cfs_rq_throttled(cfs_rq))
3211 throttle_cfs_rq(cfs_rq);
3214 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3216 struct cfs_bandwidth *cfs_b =
3217 container_of(timer, struct cfs_bandwidth, slack_timer);
3218 do_sched_cfs_slack_timer(cfs_b);
3220 return HRTIMER_NORESTART;
3223 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3225 struct cfs_bandwidth *cfs_b =
3226 container_of(timer, struct cfs_bandwidth, period_timer);
3232 now = hrtimer_cb_get_time(timer);
3233 overrun = hrtimer_forward(timer, now, cfs_b->period);
3238 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3241 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3244 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3246 raw_spin_lock_init(&cfs_b->lock);
3248 cfs_b->quota = RUNTIME_INF;
3249 cfs_b->period = ns_to_ktime(default_cfs_period());
3251 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3252 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3253 cfs_b->period_timer.function = sched_cfs_period_timer;
3254 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3255 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3258 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3260 cfs_rq->runtime_enabled = 0;
3261 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3264 /* requires cfs_b->lock, may release to reprogram timer */
3265 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3268 * The timer may be active because we're trying to set a new bandwidth
3269 * period or because we're racing with the tear-down path
3270 * (timer_active==0 becomes visible before the hrtimer call-back
3271 * terminates). In either case we ensure that it's re-programmed
3273 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3274 raw_spin_unlock(&cfs_b->lock);
3275 /* ensure cfs_b->lock is available while we wait */
3276 hrtimer_cancel(&cfs_b->period_timer);
3278 raw_spin_lock(&cfs_b->lock);
3279 /* if someone else restarted the timer then we're done */
3280 if (cfs_b->timer_active)
3284 cfs_b->timer_active = 1;
3285 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3288 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3290 hrtimer_cancel(&cfs_b->period_timer);
3291 hrtimer_cancel(&cfs_b->slack_timer);
3294 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3296 struct cfs_rq *cfs_rq;
3298 for_each_leaf_cfs_rq(rq, cfs_rq) {
3299 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3301 if (!cfs_rq->runtime_enabled)
3305 * clock_task is not advancing so we just need to make sure
3306 * there's some valid quota amount
3308 cfs_rq->runtime_remaining = cfs_b->quota;
3309 if (cfs_rq_throttled(cfs_rq))
3310 unthrottle_cfs_rq(cfs_rq);
3314 #else /* CONFIG_CFS_BANDWIDTH */
3315 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3317 return rq_clock_task(rq_of(cfs_rq));
3320 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3321 unsigned long delta_exec) {}
3322 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3323 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3324 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3326 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3331 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3336 static inline int throttled_lb_pair(struct task_group *tg,
3337 int src_cpu, int dest_cpu)
3342 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3344 #ifdef CONFIG_FAIR_GROUP_SCHED
3345 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3348 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3352 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3353 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3355 #endif /* CONFIG_CFS_BANDWIDTH */
3357 /**************************************************
3358 * CFS operations on tasks:
3361 #ifdef CONFIG_SCHED_HRTICK
3362 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3364 struct sched_entity *se = &p->se;
3365 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3367 WARN_ON(task_rq(p) != rq);
3369 if (cfs_rq->nr_running > 1) {
3370 u64 slice = sched_slice(cfs_rq, se);
3371 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3372 s64 delta = slice - ran;
3381 * Don't schedule slices shorter than 10000ns, that just
3382 * doesn't make sense. Rely on vruntime for fairness.
3385 delta = max_t(s64, 10000LL, delta);
3387 hrtick_start(rq, delta);
3392 * called from enqueue/dequeue and updates the hrtick when the
3393 * current task is from our class and nr_running is low enough
3396 static void hrtick_update(struct rq *rq)
3398 struct task_struct *curr = rq->curr;
3400 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3403 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3404 hrtick_start_fair(rq, curr);
3406 #else /* !CONFIG_SCHED_HRTICK */
3408 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3412 static inline void hrtick_update(struct rq *rq)
3418 * The enqueue_task method is called before nr_running is
3419 * increased. Here we update the fair scheduling stats and
3420 * then put the task into the rbtree:
3423 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3425 struct cfs_rq *cfs_rq;
3426 struct sched_entity *se = &p->se;
3428 for_each_sched_entity(se) {
3431 cfs_rq = cfs_rq_of(se);
3432 enqueue_entity(cfs_rq, se, flags);
3435 * end evaluation on encountering a throttled cfs_rq
3437 * note: in the case of encountering a throttled cfs_rq we will
3438 * post the final h_nr_running increment below.
3440 if (cfs_rq_throttled(cfs_rq))
3442 cfs_rq->h_nr_running++;
3444 flags = ENQUEUE_WAKEUP;
3447 for_each_sched_entity(se) {
3448 cfs_rq = cfs_rq_of(se);
3449 cfs_rq->h_nr_running++;
3451 if (cfs_rq_throttled(cfs_rq))
3454 update_cfs_shares(cfs_rq);
3455 update_entity_load_avg(se, 1);
3459 update_rq_runnable_avg(rq, rq->nr_running);
3465 static void set_next_buddy(struct sched_entity *se);
3468 * The dequeue_task method is called before nr_running is
3469 * decreased. We remove the task from the rbtree and
3470 * update the fair scheduling stats:
3472 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3474 struct cfs_rq *cfs_rq;
3475 struct sched_entity *se = &p->se;
3476 int task_sleep = flags & DEQUEUE_SLEEP;
3478 for_each_sched_entity(se) {
3479 cfs_rq = cfs_rq_of(se);
3480 dequeue_entity(cfs_rq, se, flags);
3483 * end evaluation on encountering a throttled cfs_rq
3485 * note: in the case of encountering a throttled cfs_rq we will
3486 * post the final h_nr_running decrement below.
3488 if (cfs_rq_throttled(cfs_rq))
3490 cfs_rq->h_nr_running--;
3492 /* Don't dequeue parent if it has other entities besides us */
3493 if (cfs_rq->load.weight) {
3495 * Bias pick_next to pick a task from this cfs_rq, as
3496 * p is sleeping when it is within its sched_slice.
3498 if (task_sleep && parent_entity(se))
3499 set_next_buddy(parent_entity(se));
3501 /* avoid re-evaluating load for this entity */
3502 se = parent_entity(se);
3505 flags |= DEQUEUE_SLEEP;
3508 for_each_sched_entity(se) {
3509 cfs_rq = cfs_rq_of(se);
3510 cfs_rq->h_nr_running--;
3512 if (cfs_rq_throttled(cfs_rq))
3515 update_cfs_shares(cfs_rq);
3516 update_entity_load_avg(se, 1);
3521 update_rq_runnable_avg(rq, 1);
3527 /* Used instead of source_load when we know the type == 0 */
3528 static unsigned long weighted_cpuload(const int cpu)
3530 return cpu_rq(cpu)->cfs.runnable_load_avg;
3534 * Return a low guess at the load of a migration-source cpu weighted
3535 * according to the scheduling class and "nice" value.
3537 * We want to under-estimate the load of migration sources, to
3538 * balance conservatively.
3540 static unsigned long source_load(int cpu, int type)
3542 struct rq *rq = cpu_rq(cpu);
3543 unsigned long total = weighted_cpuload(cpu);
3545 if (type == 0 || !sched_feat(LB_BIAS))
3548 return min(rq->cpu_load[type-1], total);
3552 * Return a high guess at the load of a migration-target cpu weighted
3553 * according to the scheduling class and "nice" value.
3555 static unsigned long target_load(int cpu, int type)
3557 struct rq *rq = cpu_rq(cpu);
3558 unsigned long total = weighted_cpuload(cpu);
3560 if (type == 0 || !sched_feat(LB_BIAS))
3563 return max(rq->cpu_load[type-1], total);
3566 static unsigned long power_of(int cpu)
3568 return cpu_rq(cpu)->cpu_power;
3571 static unsigned long cpu_avg_load_per_task(int cpu)
3573 struct rq *rq = cpu_rq(cpu);
3574 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3575 unsigned long load_avg = rq->cfs.runnable_load_avg;
3578 return load_avg / nr_running;
3583 static void record_wakee(struct task_struct *p)
3586 * Rough decay (wiping) for cost saving, don't worry
3587 * about the boundary, really active task won't care
3590 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3591 current->wakee_flips = 0;
3592 current->wakee_flip_decay_ts = jiffies;
3595 if (current->last_wakee != p) {
3596 current->last_wakee = p;
3597 current->wakee_flips++;
3601 static void task_waking_fair(struct task_struct *p)
3603 struct sched_entity *se = &p->se;
3604 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3607 #ifndef CONFIG_64BIT
3608 u64 min_vruntime_copy;
3611 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3613 min_vruntime = cfs_rq->min_vruntime;
3614 } while (min_vruntime != min_vruntime_copy);
3616 min_vruntime = cfs_rq->min_vruntime;
3619 se->vruntime -= min_vruntime;
3623 #ifdef CONFIG_FAIR_GROUP_SCHED
3625 * effective_load() calculates the load change as seen from the root_task_group
3627 * Adding load to a group doesn't make a group heavier, but can cause movement
3628 * of group shares between cpus. Assuming the shares were perfectly aligned one
3629 * can calculate the shift in shares.
3631 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3632 * on this @cpu and results in a total addition (subtraction) of @wg to the
3633 * total group weight.
3635 * Given a runqueue weight distribution (rw_i) we can compute a shares
3636 * distribution (s_i) using:
3638 * s_i = rw_i / \Sum rw_j (1)
3640 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3641 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3642 * shares distribution (s_i):
3644 * rw_i = { 2, 4, 1, 0 }
3645 * s_i = { 2/7, 4/7, 1/7, 0 }
3647 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3648 * task used to run on and the CPU the waker is running on), we need to
3649 * compute the effect of waking a task on either CPU and, in case of a sync
3650 * wakeup, compute the effect of the current task going to sleep.
3652 * So for a change of @wl to the local @cpu with an overall group weight change
3653 * of @wl we can compute the new shares distribution (s'_i) using:
3655 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3657 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3658 * differences in waking a task to CPU 0. The additional task changes the
3659 * weight and shares distributions like:
3661 * rw'_i = { 3, 4, 1, 0 }
3662 * s'_i = { 3/8, 4/8, 1/8, 0 }
3664 * We can then compute the difference in effective weight by using:
3666 * dw_i = S * (s'_i - s_i) (3)
3668 * Where 'S' is the group weight as seen by its parent.
3670 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3671 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3672 * 4/7) times the weight of the group.
3674 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3676 struct sched_entity *se = tg->se[cpu];
3678 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3681 for_each_sched_entity(se) {
3687 * W = @wg + \Sum rw_j
3689 W = wg + calc_tg_weight(tg, se->my_q);
3694 w = se->my_q->load.weight + wl;
3697 * wl = S * s'_i; see (2)
3700 wl = (w * tg->shares) / W;
3705 * Per the above, wl is the new se->load.weight value; since
3706 * those are clipped to [MIN_SHARES, ...) do so now. See
3707 * calc_cfs_shares().
3709 if (wl < MIN_SHARES)
3713 * wl = dw_i = S * (s'_i - s_i); see (3)
3715 wl -= se->load.weight;
3718 * Recursively apply this logic to all parent groups to compute
3719 * the final effective load change on the root group. Since
3720 * only the @tg group gets extra weight, all parent groups can
3721 * only redistribute existing shares. @wl is the shift in shares
3722 * resulting from this level per the above.
3731 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3738 static int wake_wide(struct task_struct *p)
3740 int factor = this_cpu_read(sd_llc_size);
3743 * Yeah, it's the switching-frequency, could means many wakee or
3744 * rapidly switch, use factor here will just help to automatically
3745 * adjust the loose-degree, so bigger node will lead to more pull.
3747 if (p->wakee_flips > factor) {
3749 * wakee is somewhat hot, it needs certain amount of cpu
3750 * resource, so if waker is far more hot, prefer to leave
3753 if (current->wakee_flips > (factor * p->wakee_flips))
3760 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3762 s64 this_load, load;
3763 int idx, this_cpu, prev_cpu;
3764 unsigned long tl_per_task;
3765 struct task_group *tg;
3766 unsigned long weight;
3770 * If we wake multiple tasks be careful to not bounce
3771 * ourselves around too much.
3777 this_cpu = smp_processor_id();
3778 prev_cpu = task_cpu(p);
3779 load = source_load(prev_cpu, idx);
3780 this_load = target_load(this_cpu, idx);
3783 * If sync wakeup then subtract the (maximum possible)
3784 * effect of the currently running task from the load
3785 * of the current CPU:
3788 tg = task_group(current);
3789 weight = current->se.load.weight;
3791 this_load += effective_load(tg, this_cpu, -weight, -weight);
3792 load += effective_load(tg, prev_cpu, 0, -weight);
3796 weight = p->se.load.weight;
3799 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3800 * due to the sync cause above having dropped this_load to 0, we'll
3801 * always have an imbalance, but there's really nothing you can do
3802 * about that, so that's good too.
3804 * Otherwise check if either cpus are near enough in load to allow this
3805 * task to be woken on this_cpu.
3807 if (this_load > 0) {
3808 s64 this_eff_load, prev_eff_load;
3810 this_eff_load = 100;
3811 this_eff_load *= power_of(prev_cpu);
3812 this_eff_load *= this_load +
3813 effective_load(tg, this_cpu, weight, weight);
3815 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3816 prev_eff_load *= power_of(this_cpu);
3817 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3819 balanced = this_eff_load <= prev_eff_load;
3824 * If the currently running task will sleep within
3825 * a reasonable amount of time then attract this newly
3828 if (sync && balanced)
3831 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3832 tl_per_task = cpu_avg_load_per_task(this_cpu);
3835 (this_load <= load &&
3836 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3838 * This domain has SD_WAKE_AFFINE and
3839 * p is cache cold in this domain, and
3840 * there is no bad imbalance.
3842 schedstat_inc(sd, ttwu_move_affine);
3843 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3851 * find_idlest_group finds and returns the least busy CPU group within the
3854 static struct sched_group *
3855 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3856 int this_cpu, int load_idx)
3858 struct sched_group *idlest = NULL, *group = sd->groups;
3859 unsigned long min_load = ULONG_MAX, this_load = 0;
3860 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3863 unsigned long load, avg_load;
3867 /* Skip over this group if it has no CPUs allowed */
3868 if (!cpumask_intersects(sched_group_cpus(group),
3869 tsk_cpus_allowed(p)))
3872 local_group = cpumask_test_cpu(this_cpu,
3873 sched_group_cpus(group));
3875 /* Tally up the load of all CPUs in the group */
3878 for_each_cpu(i, sched_group_cpus(group)) {
3879 /* Bias balancing toward cpus of our domain */
3881 load = source_load(i, load_idx);
3883 load = target_load(i, load_idx);
3888 /* Adjust by relative CPU power of the group */
3889 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3892 this_load = avg_load;
3893 } else if (avg_load < min_load) {
3894 min_load = avg_load;
3897 } while (group = group->next, group != sd->groups);
3899 if (!idlest || 100*this_load < imbalance*min_load)
3905 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3908 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3910 unsigned long load, min_load = ULONG_MAX;
3914 /* Traverse only the allowed CPUs */
3915 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3916 load = weighted_cpuload(i);
3918 if (load < min_load || (load == min_load && i == this_cpu)) {
3928 * Try and locate an idle CPU in the sched_domain.
3930 static int select_idle_sibling(struct task_struct *p, int target)
3932 struct sched_domain *sd;
3933 struct sched_group *sg;
3934 int i = task_cpu(p);
3936 if (idle_cpu(target))
3940 * If the prevous cpu is cache affine and idle, don't be stupid.
3942 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3946 * Otherwise, iterate the domains and find an elegible idle cpu.
3948 sd = rcu_dereference(per_cpu(sd_llc, target));
3949 for_each_lower_domain(sd) {
3952 if (!cpumask_intersects(sched_group_cpus(sg),
3953 tsk_cpus_allowed(p)))
3956 for_each_cpu(i, sched_group_cpus(sg)) {
3957 if (i == target || !idle_cpu(i))
3961 target = cpumask_first_and(sched_group_cpus(sg),
3962 tsk_cpus_allowed(p));
3966 } while (sg != sd->groups);
3973 * sched_balance_self: balance the current task (running on cpu) in domains
3974 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3977 * Balance, ie. select the least loaded group.
3979 * Returns the target CPU number, or the same CPU if no balancing is needed.
3981 * preempt must be disabled.
3984 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
3986 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3987 int cpu = smp_processor_id();
3989 int want_affine = 0;
3990 int sync = wake_flags & WF_SYNC;
3992 if (p->nr_cpus_allowed == 1)
3995 if (sd_flag & SD_BALANCE_WAKE) {
3996 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4002 for_each_domain(cpu, tmp) {
4003 if (!(tmp->flags & SD_LOAD_BALANCE))
4007 * If both cpu and prev_cpu are part of this domain,
4008 * cpu is a valid SD_WAKE_AFFINE target.
4010 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4011 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4016 if (tmp->flags & sd_flag)
4021 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4024 new_cpu = select_idle_sibling(p, prev_cpu);
4029 int load_idx = sd->forkexec_idx;
4030 struct sched_group *group;
4033 if (!(sd->flags & sd_flag)) {
4038 if (sd_flag & SD_BALANCE_WAKE)
4039 load_idx = sd->wake_idx;
4041 group = find_idlest_group(sd, p, cpu, load_idx);
4047 new_cpu = find_idlest_cpu(group, p, cpu);
4048 if (new_cpu == -1 || new_cpu == cpu) {
4049 /* Now try balancing at a lower domain level of cpu */
4054 /* Now try balancing at a lower domain level of new_cpu */
4056 weight = sd->span_weight;
4058 for_each_domain(cpu, tmp) {
4059 if (weight <= tmp->span_weight)
4061 if (tmp->flags & sd_flag)
4064 /* while loop will break here if sd == NULL */
4073 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4074 * cfs_rq_of(p) references at time of call are still valid and identify the
4075 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4076 * other assumptions, including the state of rq->lock, should be made.
4079 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4081 struct sched_entity *se = &p->se;
4082 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4085 * Load tracking: accumulate removed load so that it can be processed
4086 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4087 * to blocked load iff they have a positive decay-count. It can never
4088 * be negative here since on-rq tasks have decay-count == 0.
4090 if (se->avg.decay_count) {
4091 se->avg.decay_count = -__synchronize_entity_decay(se);
4092 atomic_long_add(se->avg.load_avg_contrib,
4093 &cfs_rq->removed_load);
4096 #endif /* CONFIG_SMP */
4098 static unsigned long
4099 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4101 unsigned long gran = sysctl_sched_wakeup_granularity;
4104 * Since its curr running now, convert the gran from real-time
4105 * to virtual-time in his units.
4107 * By using 'se' instead of 'curr' we penalize light tasks, so
4108 * they get preempted easier. That is, if 'se' < 'curr' then
4109 * the resulting gran will be larger, therefore penalizing the
4110 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4111 * be smaller, again penalizing the lighter task.
4113 * This is especially important for buddies when the leftmost
4114 * task is higher priority than the buddy.
4116 return calc_delta_fair(gran, se);
4120 * Should 'se' preempt 'curr'.
4134 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4136 s64 gran, vdiff = curr->vruntime - se->vruntime;
4141 gran = wakeup_gran(curr, se);
4148 static void set_last_buddy(struct sched_entity *se)
4150 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4153 for_each_sched_entity(se)
4154 cfs_rq_of(se)->last = se;
4157 static void set_next_buddy(struct sched_entity *se)
4159 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4162 for_each_sched_entity(se)
4163 cfs_rq_of(se)->next = se;
4166 static void set_skip_buddy(struct sched_entity *se)
4168 for_each_sched_entity(se)
4169 cfs_rq_of(se)->skip = se;
4173 * Preempt the current task with a newly woken task if needed:
4175 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4177 struct task_struct *curr = rq->curr;
4178 struct sched_entity *se = &curr->se, *pse = &p->se;
4179 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4180 int scale = cfs_rq->nr_running >= sched_nr_latency;
4181 int next_buddy_marked = 0;
4183 if (unlikely(se == pse))
4187 * This is possible from callers such as move_task(), in which we
4188 * unconditionally check_prempt_curr() after an enqueue (which may have
4189 * lead to a throttle). This both saves work and prevents false
4190 * next-buddy nomination below.
4192 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4195 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4196 set_next_buddy(pse);
4197 next_buddy_marked = 1;
4201 * We can come here with TIF_NEED_RESCHED already set from new task
4204 * Note: this also catches the edge-case of curr being in a throttled
4205 * group (e.g. via set_curr_task), since update_curr() (in the
4206 * enqueue of curr) will have resulted in resched being set. This
4207 * prevents us from potentially nominating it as a false LAST_BUDDY
4210 if (test_tsk_need_resched(curr))
4213 /* Idle tasks are by definition preempted by non-idle tasks. */
4214 if (unlikely(curr->policy == SCHED_IDLE) &&
4215 likely(p->policy != SCHED_IDLE))
4219 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4220 * is driven by the tick):
4222 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4225 find_matching_se(&se, &pse);
4226 update_curr(cfs_rq_of(se));
4228 if (wakeup_preempt_entity(se, pse) == 1) {
4230 * Bias pick_next to pick the sched entity that is
4231 * triggering this preemption.
4233 if (!next_buddy_marked)
4234 set_next_buddy(pse);
4243 * Only set the backward buddy when the current task is still
4244 * on the rq. This can happen when a wakeup gets interleaved
4245 * with schedule on the ->pre_schedule() or idle_balance()
4246 * point, either of which can * drop the rq lock.
4248 * Also, during early boot the idle thread is in the fair class,
4249 * for obvious reasons its a bad idea to schedule back to it.
4251 if (unlikely(!se->on_rq || curr == rq->idle))
4254 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4258 static struct task_struct *pick_next_task_fair(struct rq *rq)
4260 struct task_struct *p;
4261 struct cfs_rq *cfs_rq = &rq->cfs;
4262 struct sched_entity *se;
4264 if (!cfs_rq->nr_running)
4268 se = pick_next_entity(cfs_rq);
4269 set_next_entity(cfs_rq, se);
4270 cfs_rq = group_cfs_rq(se);
4274 if (hrtick_enabled(rq))
4275 hrtick_start_fair(rq, p);
4281 * Account for a descheduled task:
4283 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4285 struct sched_entity *se = &prev->se;
4286 struct cfs_rq *cfs_rq;
4288 for_each_sched_entity(se) {
4289 cfs_rq = cfs_rq_of(se);
4290 put_prev_entity(cfs_rq, se);
4295 * sched_yield() is very simple
4297 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4299 static void yield_task_fair(struct rq *rq)
4301 struct task_struct *curr = rq->curr;
4302 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4303 struct sched_entity *se = &curr->se;
4306 * Are we the only task in the tree?
4308 if (unlikely(rq->nr_running == 1))
4311 clear_buddies(cfs_rq, se);
4313 if (curr->policy != SCHED_BATCH) {
4314 update_rq_clock(rq);
4316 * Update run-time statistics of the 'current'.
4318 update_curr(cfs_rq);
4320 * Tell update_rq_clock() that we've just updated,
4321 * so we don't do microscopic update in schedule()
4322 * and double the fastpath cost.
4324 rq->skip_clock_update = 1;
4330 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4332 struct sched_entity *se = &p->se;
4334 /* throttled hierarchies are not runnable */
4335 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4338 /* Tell the scheduler that we'd really like pse to run next. */
4341 yield_task_fair(rq);
4347 /**************************************************
4348 * Fair scheduling class load-balancing methods.
4352 * The purpose of load-balancing is to achieve the same basic fairness the
4353 * per-cpu scheduler provides, namely provide a proportional amount of compute
4354 * time to each task. This is expressed in the following equation:
4356 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4358 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4359 * W_i,0 is defined as:
4361 * W_i,0 = \Sum_j w_i,j (2)
4363 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4364 * is derived from the nice value as per prio_to_weight[].
4366 * The weight average is an exponential decay average of the instantaneous
4369 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4371 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4372 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4373 * can also include other factors [XXX].
4375 * To achieve this balance we define a measure of imbalance which follows
4376 * directly from (1):
4378 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4380 * We them move tasks around to minimize the imbalance. In the continuous
4381 * function space it is obvious this converges, in the discrete case we get
4382 * a few fun cases generally called infeasible weight scenarios.
4385 * - infeasible weights;
4386 * - local vs global optima in the discrete case. ]
4391 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4392 * for all i,j solution, we create a tree of cpus that follows the hardware
4393 * topology where each level pairs two lower groups (or better). This results
4394 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4395 * tree to only the first of the previous level and we decrease the frequency
4396 * of load-balance at each level inv. proportional to the number of cpus in
4402 * \Sum { --- * --- * 2^i } = O(n) (5)
4404 * `- size of each group
4405 * | | `- number of cpus doing load-balance
4407 * `- sum over all levels
4409 * Coupled with a limit on how many tasks we can migrate every balance pass,
4410 * this makes (5) the runtime complexity of the balancer.
4412 * An important property here is that each CPU is still (indirectly) connected
4413 * to every other cpu in at most O(log n) steps:
4415 * The adjacency matrix of the resulting graph is given by:
4418 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4421 * And you'll find that:
4423 * A^(log_2 n)_i,j != 0 for all i,j (7)
4425 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4426 * The task movement gives a factor of O(m), giving a convergence complexity
4429 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4434 * In order to avoid CPUs going idle while there's still work to do, new idle
4435 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4436 * tree itself instead of relying on other CPUs to bring it work.
4438 * This adds some complexity to both (5) and (8) but it reduces the total idle
4446 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4449 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4454 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4456 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4458 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4461 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4462 * rewrite all of this once again.]
4465 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4467 #define LBF_ALL_PINNED 0x01
4468 #define LBF_NEED_BREAK 0x02
4469 #define LBF_DST_PINNED 0x04
4470 #define LBF_SOME_PINNED 0x08
4473 struct sched_domain *sd;
4481 struct cpumask *dst_grpmask;
4483 enum cpu_idle_type idle;
4485 /* The set of CPUs under consideration for load-balancing */
4486 struct cpumask *cpus;
4491 unsigned int loop_break;
4492 unsigned int loop_max;
4496 * move_task - move a task from one runqueue to another runqueue.
4497 * Both runqueues must be locked.
4499 static void move_task(struct task_struct *p, struct lb_env *env)
4501 deactivate_task(env->src_rq, p, 0);
4502 set_task_cpu(p, env->dst_cpu);
4503 activate_task(env->dst_rq, p, 0);
4504 check_preempt_curr(env->dst_rq, p, 0);
4505 #ifdef CONFIG_NUMA_BALANCING
4506 if (p->numa_preferred_nid != -1) {
4507 int src_nid = cpu_to_node(env->src_cpu);
4508 int dst_nid = cpu_to_node(env->dst_cpu);
4511 * If the load balancer has moved the task then limit
4512 * migrations from taking place in the short term in
4513 * case this is a short-lived migration.
4515 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4516 p->numa_migrate_seq = 0;
4522 * Is this task likely cache-hot:
4525 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4529 if (p->sched_class != &fair_sched_class)
4532 if (unlikely(p->policy == SCHED_IDLE))
4536 * Buddy candidates are cache hot:
4538 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4539 (&p->se == cfs_rq_of(&p->se)->next ||
4540 &p->se == cfs_rq_of(&p->se)->last))
4543 if (sysctl_sched_migration_cost == -1)
4545 if (sysctl_sched_migration_cost == 0)
4548 delta = now - p->se.exec_start;
4550 return delta < (s64)sysctl_sched_migration_cost;
4553 #ifdef CONFIG_NUMA_BALANCING
4554 /* Returns true if the destination node has incurred more faults */
4555 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4557 int src_nid, dst_nid;
4559 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4560 !(env->sd->flags & SD_NUMA)) {
4564 src_nid = cpu_to_node(env->src_cpu);
4565 dst_nid = cpu_to_node(env->dst_cpu);
4567 if (src_nid == dst_nid ||
4568 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4571 if (dst_nid == p->numa_preferred_nid ||
4572 task_faults(p, dst_nid) > task_faults(p, src_nid))
4579 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4581 int src_nid, dst_nid;
4583 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4586 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4589 src_nid = cpu_to_node(env->src_cpu);
4590 dst_nid = cpu_to_node(env->dst_cpu);
4592 if (src_nid == dst_nid ||
4593 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4596 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4603 static inline bool migrate_improves_locality(struct task_struct *p,
4609 static inline bool migrate_degrades_locality(struct task_struct *p,
4617 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4620 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4622 int tsk_cache_hot = 0;
4624 * We do not migrate tasks that are:
4625 * 1) throttled_lb_pair, or
4626 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4627 * 3) running (obviously), or
4628 * 4) are cache-hot on their current CPU.
4630 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4633 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4636 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4638 env->flags |= LBF_SOME_PINNED;
4641 * Remember if this task can be migrated to any other cpu in
4642 * our sched_group. We may want to revisit it if we couldn't
4643 * meet load balance goals by pulling other tasks on src_cpu.
4645 * Also avoid computing new_dst_cpu if we have already computed
4646 * one in current iteration.
4648 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4651 /* Prevent to re-select dst_cpu via env's cpus */
4652 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4653 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4654 env->flags |= LBF_DST_PINNED;
4655 env->new_dst_cpu = cpu;
4663 /* Record that we found atleast one task that could run on dst_cpu */
4664 env->flags &= ~LBF_ALL_PINNED;
4666 if (task_running(env->src_rq, p)) {
4667 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4672 * Aggressive migration if:
4673 * 1) destination numa is preferred
4674 * 2) task is cache cold, or
4675 * 3) too many balance attempts have failed.
4677 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4679 tsk_cache_hot = migrate_degrades_locality(p, env);
4681 if (migrate_improves_locality(p, env)) {
4682 #ifdef CONFIG_SCHEDSTATS
4683 if (tsk_cache_hot) {
4684 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4685 schedstat_inc(p, se.statistics.nr_forced_migrations);
4691 if (!tsk_cache_hot ||
4692 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4694 if (tsk_cache_hot) {
4695 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4696 schedstat_inc(p, se.statistics.nr_forced_migrations);
4702 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4707 * move_one_task tries to move exactly one task from busiest to this_rq, as
4708 * part of active balancing operations within "domain".
4709 * Returns 1 if successful and 0 otherwise.
4711 * Called with both runqueues locked.
4713 static int move_one_task(struct lb_env *env)
4715 struct task_struct *p, *n;
4717 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4718 if (!can_migrate_task(p, env))
4723 * Right now, this is only the second place move_task()
4724 * is called, so we can safely collect move_task()
4725 * stats here rather than inside move_task().
4727 schedstat_inc(env->sd, lb_gained[env->idle]);
4733 static const unsigned int sched_nr_migrate_break = 32;
4736 * move_tasks tries to move up to imbalance weighted load from busiest to
4737 * this_rq, as part of a balancing operation within domain "sd".
4738 * Returns 1 if successful and 0 otherwise.
4740 * Called with both runqueues locked.
4742 static int move_tasks(struct lb_env *env)
4744 struct list_head *tasks = &env->src_rq->cfs_tasks;
4745 struct task_struct *p;
4749 if (env->imbalance <= 0)
4752 while (!list_empty(tasks)) {
4753 p = list_first_entry(tasks, struct task_struct, se.group_node);
4756 /* We've more or less seen every task there is, call it quits */
4757 if (env->loop > env->loop_max)
4760 /* take a breather every nr_migrate tasks */
4761 if (env->loop > env->loop_break) {
4762 env->loop_break += sched_nr_migrate_break;
4763 env->flags |= LBF_NEED_BREAK;
4767 if (!can_migrate_task(p, env))
4770 load = task_h_load(p);
4772 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4775 if ((load / 2) > env->imbalance)
4780 env->imbalance -= load;
4782 #ifdef CONFIG_PREEMPT
4784 * NEWIDLE balancing is a source of latency, so preemptible
4785 * kernels will stop after the first task is pulled to minimize
4786 * the critical section.
4788 if (env->idle == CPU_NEWLY_IDLE)
4793 * We only want to steal up to the prescribed amount of
4796 if (env->imbalance <= 0)
4801 list_move_tail(&p->se.group_node, tasks);
4805 * Right now, this is one of only two places move_task() is called,
4806 * so we can safely collect move_task() stats here rather than
4807 * inside move_task().
4809 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4814 #ifdef CONFIG_FAIR_GROUP_SCHED
4816 * update tg->load_weight by folding this cpu's load_avg
4818 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4820 struct sched_entity *se = tg->se[cpu];
4821 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4823 /* throttled entities do not contribute to load */
4824 if (throttled_hierarchy(cfs_rq))
4827 update_cfs_rq_blocked_load(cfs_rq, 1);
4830 update_entity_load_avg(se, 1);
4832 * We pivot on our runnable average having decayed to zero for
4833 * list removal. This generally implies that all our children
4834 * have also been removed (modulo rounding error or bandwidth
4835 * control); however, such cases are rare and we can fix these
4838 * TODO: fix up out-of-order children on enqueue.
4840 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4841 list_del_leaf_cfs_rq(cfs_rq);
4843 struct rq *rq = rq_of(cfs_rq);
4844 update_rq_runnable_avg(rq, rq->nr_running);
4848 static void update_blocked_averages(int cpu)
4850 struct rq *rq = cpu_rq(cpu);
4851 struct cfs_rq *cfs_rq;
4852 unsigned long flags;
4854 raw_spin_lock_irqsave(&rq->lock, flags);
4855 update_rq_clock(rq);
4857 * Iterates the task_group tree in a bottom up fashion, see
4858 * list_add_leaf_cfs_rq() for details.
4860 for_each_leaf_cfs_rq(rq, cfs_rq) {
4862 * Note: We may want to consider periodically releasing
4863 * rq->lock about these updates so that creating many task
4864 * groups does not result in continually extending hold time.
4866 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4869 raw_spin_unlock_irqrestore(&rq->lock, flags);
4873 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4874 * This needs to be done in a top-down fashion because the load of a child
4875 * group is a fraction of its parents load.
4877 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4879 struct rq *rq = rq_of(cfs_rq);
4880 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4881 unsigned long now = jiffies;
4884 if (cfs_rq->last_h_load_update == now)
4887 cfs_rq->h_load_next = NULL;
4888 for_each_sched_entity(se) {
4889 cfs_rq = cfs_rq_of(se);
4890 cfs_rq->h_load_next = se;
4891 if (cfs_rq->last_h_load_update == now)
4896 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4897 cfs_rq->last_h_load_update = now;
4900 while ((se = cfs_rq->h_load_next) != NULL) {
4901 load = cfs_rq->h_load;
4902 load = div64_ul(load * se->avg.load_avg_contrib,
4903 cfs_rq->runnable_load_avg + 1);
4904 cfs_rq = group_cfs_rq(se);
4905 cfs_rq->h_load = load;
4906 cfs_rq->last_h_load_update = now;
4910 static unsigned long task_h_load(struct task_struct *p)
4912 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4914 update_cfs_rq_h_load(cfs_rq);
4915 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4916 cfs_rq->runnable_load_avg + 1);
4919 static inline void update_blocked_averages(int cpu)
4923 static unsigned long task_h_load(struct task_struct *p)
4925 return p->se.avg.load_avg_contrib;
4929 /********** Helpers for find_busiest_group ************************/
4931 * sg_lb_stats - stats of a sched_group required for load_balancing
4933 struct sg_lb_stats {
4934 unsigned long avg_load; /*Avg load across the CPUs of the group */
4935 unsigned long group_load; /* Total load over the CPUs of the group */
4936 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4937 unsigned long load_per_task;
4938 unsigned long group_power;
4939 unsigned int sum_nr_running; /* Nr tasks running in the group */
4940 unsigned int group_capacity;
4941 unsigned int idle_cpus;
4942 unsigned int group_weight;
4943 int group_imb; /* Is there an imbalance in the group ? */
4944 int group_has_capacity; /* Is there extra capacity in the group? */
4948 * sd_lb_stats - Structure to store the statistics of a sched_domain
4949 * during load balancing.
4951 struct sd_lb_stats {
4952 struct sched_group *busiest; /* Busiest group in this sd */
4953 struct sched_group *local; /* Local group in this sd */
4954 unsigned long total_load; /* Total load of all groups in sd */
4955 unsigned long total_pwr; /* Total power of all groups in sd */
4956 unsigned long avg_load; /* Average load across all groups in sd */
4958 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4959 struct sg_lb_stats local_stat; /* Statistics of the local group */
4962 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4965 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4966 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4967 * We must however clear busiest_stat::avg_load because
4968 * update_sd_pick_busiest() reads this before assignment.
4970 *sds = (struct sd_lb_stats){
4982 * get_sd_load_idx - Obtain the load index for a given sched domain.
4983 * @sd: The sched_domain whose load_idx is to be obtained.
4984 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4986 * Return: The load index.
4988 static inline int get_sd_load_idx(struct sched_domain *sd,
4989 enum cpu_idle_type idle)
4995 load_idx = sd->busy_idx;
4998 case CPU_NEWLY_IDLE:
4999 load_idx = sd->newidle_idx;
5002 load_idx = sd->idle_idx;
5009 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5011 return SCHED_POWER_SCALE;
5014 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5016 return default_scale_freq_power(sd, cpu);
5019 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5021 unsigned long weight = sd->span_weight;
5022 unsigned long smt_gain = sd->smt_gain;
5029 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5031 return default_scale_smt_power(sd, cpu);
5034 static unsigned long scale_rt_power(int cpu)
5036 struct rq *rq = cpu_rq(cpu);
5037 u64 total, available, age_stamp, avg;
5040 * Since we're reading these variables without serialization make sure
5041 * we read them once before doing sanity checks on them.
5043 age_stamp = ACCESS_ONCE(rq->age_stamp);
5044 avg = ACCESS_ONCE(rq->rt_avg);
5046 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5048 if (unlikely(total < avg)) {
5049 /* Ensures that power won't end up being negative */
5052 available = total - avg;
5055 if (unlikely((s64)total < SCHED_POWER_SCALE))
5056 total = SCHED_POWER_SCALE;
5058 total >>= SCHED_POWER_SHIFT;
5060 return div_u64(available, total);
5063 static void update_cpu_power(struct sched_domain *sd, int cpu)
5065 unsigned long weight = sd->span_weight;
5066 unsigned long power = SCHED_POWER_SCALE;
5067 struct sched_group *sdg = sd->groups;
5069 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5070 if (sched_feat(ARCH_POWER))
5071 power *= arch_scale_smt_power(sd, cpu);
5073 power *= default_scale_smt_power(sd, cpu);
5075 power >>= SCHED_POWER_SHIFT;
5078 sdg->sgp->power_orig = power;
5080 if (sched_feat(ARCH_POWER))
5081 power *= arch_scale_freq_power(sd, cpu);
5083 power *= default_scale_freq_power(sd, cpu);
5085 power >>= SCHED_POWER_SHIFT;
5087 power *= scale_rt_power(cpu);
5088 power >>= SCHED_POWER_SHIFT;
5093 cpu_rq(cpu)->cpu_power = power;
5094 sdg->sgp->power = power;
5097 void update_group_power(struct sched_domain *sd, int cpu)
5099 struct sched_domain *child = sd->child;
5100 struct sched_group *group, *sdg = sd->groups;
5101 unsigned long power, power_orig;
5102 unsigned long interval;
5104 interval = msecs_to_jiffies(sd->balance_interval);
5105 interval = clamp(interval, 1UL, max_load_balance_interval);
5106 sdg->sgp->next_update = jiffies + interval;
5109 update_cpu_power(sd, cpu);
5113 power_orig = power = 0;
5115 if (child->flags & SD_OVERLAP) {
5117 * SD_OVERLAP domains cannot assume that child groups
5118 * span the current group.
5121 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5122 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5124 power_orig += sg->sgp->power_orig;
5125 power += sg->sgp->power;
5129 * !SD_OVERLAP domains can assume that child groups
5130 * span the current group.
5133 group = child->groups;
5135 power_orig += group->sgp->power_orig;
5136 power += group->sgp->power;
5137 group = group->next;
5138 } while (group != child->groups);
5141 sdg->sgp->power_orig = power_orig;
5142 sdg->sgp->power = power;
5146 * Try and fix up capacity for tiny siblings, this is needed when
5147 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5148 * which on its own isn't powerful enough.
5150 * See update_sd_pick_busiest() and check_asym_packing().
5153 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5156 * Only siblings can have significantly less than SCHED_POWER_SCALE
5158 if (!(sd->flags & SD_SHARE_CPUPOWER))
5162 * If ~90% of the cpu_power is still there, we're good.
5164 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5171 * Group imbalance indicates (and tries to solve) the problem where balancing
5172 * groups is inadequate due to tsk_cpus_allowed() constraints.
5174 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5175 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5178 * { 0 1 2 3 } { 4 5 6 7 }
5181 * If we were to balance group-wise we'd place two tasks in the first group and
5182 * two tasks in the second group. Clearly this is undesired as it will overload
5183 * cpu 3 and leave one of the cpus in the second group unused.
5185 * The current solution to this issue is detecting the skew in the first group
5186 * by noticing the lower domain failed to reach balance and had difficulty
5187 * moving tasks due to affinity constraints.
5189 * When this is so detected; this group becomes a candidate for busiest; see
5190 * update_sd_pick_busiest(). And calculcate_imbalance() and
5191 * find_busiest_group() avoid some of the usual balance conditions to allow it
5192 * to create an effective group imbalance.
5194 * This is a somewhat tricky proposition since the next run might not find the
5195 * group imbalance and decide the groups need to be balanced again. A most
5196 * subtle and fragile situation.
5199 static inline int sg_imbalanced(struct sched_group *group)
5201 return group->sgp->imbalance;
5205 * Compute the group capacity.
5207 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5208 * first dividing out the smt factor and computing the actual number of cores
5209 * and limit power unit capacity with that.
5211 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5213 unsigned int capacity, smt, cpus;
5214 unsigned int power, power_orig;
5216 power = group->sgp->power;
5217 power_orig = group->sgp->power_orig;
5218 cpus = group->group_weight;
5220 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5221 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5222 capacity = cpus / smt; /* cores */
5224 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5226 capacity = fix_small_capacity(env->sd, group);
5232 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5233 * @env: The load balancing environment.
5234 * @group: sched_group whose statistics are to be updated.
5235 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5236 * @local_group: Does group contain this_cpu.
5237 * @sgs: variable to hold the statistics for this group.
5239 static inline void update_sg_lb_stats(struct lb_env *env,
5240 struct sched_group *group, int load_idx,
5241 int local_group, struct sg_lb_stats *sgs)
5243 unsigned long nr_running;
5247 memset(sgs, 0, sizeof(*sgs));
5249 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5250 struct rq *rq = cpu_rq(i);
5252 nr_running = rq->nr_running;
5254 /* Bias balancing toward cpus of our domain */
5256 load = target_load(i, load_idx);
5258 load = source_load(i, load_idx);
5260 sgs->group_load += load;
5261 sgs->sum_nr_running += nr_running;
5262 sgs->sum_weighted_load += weighted_cpuload(i);
5267 /* Adjust by relative CPU power of the group */
5268 sgs->group_power = group->sgp->power;
5269 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5271 if (sgs->sum_nr_running)
5272 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5274 sgs->group_weight = group->group_weight;
5276 sgs->group_imb = sg_imbalanced(group);
5277 sgs->group_capacity = sg_capacity(env, group);
5279 if (sgs->group_capacity > sgs->sum_nr_running)
5280 sgs->group_has_capacity = 1;
5284 * update_sd_pick_busiest - return 1 on busiest group
5285 * @env: The load balancing environment.
5286 * @sds: sched_domain statistics
5287 * @sg: sched_group candidate to be checked for being the busiest
5288 * @sgs: sched_group statistics
5290 * Determine if @sg is a busier group than the previously selected
5293 * Return: %true if @sg is a busier group than the previously selected
5294 * busiest group. %false otherwise.
5296 static bool update_sd_pick_busiest(struct lb_env *env,
5297 struct sd_lb_stats *sds,
5298 struct sched_group *sg,
5299 struct sg_lb_stats *sgs)
5301 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5304 if (sgs->sum_nr_running > sgs->group_capacity)
5311 * ASYM_PACKING needs to move all the work to the lowest
5312 * numbered CPUs in the group, therefore mark all groups
5313 * higher than ourself as busy.
5315 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5316 env->dst_cpu < group_first_cpu(sg)) {
5320 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5328 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5329 * @env: The load balancing environment.
5330 * @balance: Should we balance.
5331 * @sds: variable to hold the statistics for this sched_domain.
5333 static inline void update_sd_lb_stats(struct lb_env *env,
5334 struct sd_lb_stats *sds)
5336 struct sched_domain *child = env->sd->child;
5337 struct sched_group *sg = env->sd->groups;
5338 struct sg_lb_stats tmp_sgs;
5339 int load_idx, prefer_sibling = 0;
5341 if (child && child->flags & SD_PREFER_SIBLING)
5344 load_idx = get_sd_load_idx(env->sd, env->idle);
5347 struct sg_lb_stats *sgs = &tmp_sgs;
5350 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5353 sgs = &sds->local_stat;
5355 if (env->idle != CPU_NEWLY_IDLE ||
5356 time_after_eq(jiffies, sg->sgp->next_update))
5357 update_group_power(env->sd, env->dst_cpu);
5360 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5366 * In case the child domain prefers tasks go to siblings
5367 * first, lower the sg capacity to one so that we'll try
5368 * and move all the excess tasks away. We lower the capacity
5369 * of a group only if the local group has the capacity to fit
5370 * these excess tasks, i.e. nr_running < group_capacity. The
5371 * extra check prevents the case where you always pull from the
5372 * heaviest group when it is already under-utilized (possible
5373 * with a large weight task outweighs the tasks on the system).
5375 if (prefer_sibling && sds->local &&
5376 sds->local_stat.group_has_capacity)
5377 sgs->group_capacity = min(sgs->group_capacity, 1U);
5379 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5381 sds->busiest_stat = *sgs;
5385 /* Now, start updating sd_lb_stats */
5386 sds->total_load += sgs->group_load;
5387 sds->total_pwr += sgs->group_power;
5390 } while (sg != env->sd->groups);
5394 * check_asym_packing - Check to see if the group is packed into the
5397 * This is primarily intended to used at the sibling level. Some
5398 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5399 * case of POWER7, it can move to lower SMT modes only when higher
5400 * threads are idle. When in lower SMT modes, the threads will
5401 * perform better since they share less core resources. Hence when we
5402 * have idle threads, we want them to be the higher ones.
5404 * This packing function is run on idle threads. It checks to see if
5405 * the busiest CPU in this domain (core in the P7 case) has a higher
5406 * CPU number than the packing function is being run on. Here we are
5407 * assuming lower CPU number will be equivalent to lower a SMT thread
5410 * Return: 1 when packing is required and a task should be moved to
5411 * this CPU. The amount of the imbalance is returned in *imbalance.
5413 * @env: The load balancing environment.
5414 * @sds: Statistics of the sched_domain which is to be packed
5416 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5420 if (!(env->sd->flags & SD_ASYM_PACKING))
5426 busiest_cpu = group_first_cpu(sds->busiest);
5427 if (env->dst_cpu > busiest_cpu)
5430 env->imbalance = DIV_ROUND_CLOSEST(
5431 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5438 * fix_small_imbalance - Calculate the minor imbalance that exists
5439 * amongst the groups of a sched_domain, during
5441 * @env: The load balancing environment.
5442 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5445 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5447 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5448 unsigned int imbn = 2;
5449 unsigned long scaled_busy_load_per_task;
5450 struct sg_lb_stats *local, *busiest;
5452 local = &sds->local_stat;
5453 busiest = &sds->busiest_stat;
5455 if (!local->sum_nr_running)
5456 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5457 else if (busiest->load_per_task > local->load_per_task)
5460 scaled_busy_load_per_task =
5461 (busiest->load_per_task * SCHED_POWER_SCALE) /
5462 busiest->group_power;
5464 if (busiest->avg_load + scaled_busy_load_per_task >=
5465 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5466 env->imbalance = busiest->load_per_task;
5471 * OK, we don't have enough imbalance to justify moving tasks,
5472 * however we may be able to increase total CPU power used by
5476 pwr_now += busiest->group_power *
5477 min(busiest->load_per_task, busiest->avg_load);
5478 pwr_now += local->group_power *
5479 min(local->load_per_task, local->avg_load);
5480 pwr_now /= SCHED_POWER_SCALE;
5482 /* Amount of load we'd subtract */
5483 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5484 busiest->group_power;
5485 if (busiest->avg_load > tmp) {
5486 pwr_move += busiest->group_power *
5487 min(busiest->load_per_task,
5488 busiest->avg_load - tmp);
5491 /* Amount of load we'd add */
5492 if (busiest->avg_load * busiest->group_power <
5493 busiest->load_per_task * SCHED_POWER_SCALE) {
5494 tmp = (busiest->avg_load * busiest->group_power) /
5497 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5500 pwr_move += local->group_power *
5501 min(local->load_per_task, local->avg_load + tmp);
5502 pwr_move /= SCHED_POWER_SCALE;
5504 /* Move if we gain throughput */
5505 if (pwr_move > pwr_now)
5506 env->imbalance = busiest->load_per_task;
5510 * calculate_imbalance - Calculate the amount of imbalance present within the
5511 * groups of a given sched_domain during load balance.
5512 * @env: load balance environment
5513 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5515 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5517 unsigned long max_pull, load_above_capacity = ~0UL;
5518 struct sg_lb_stats *local, *busiest;
5520 local = &sds->local_stat;
5521 busiest = &sds->busiest_stat;
5523 if (busiest->group_imb) {
5525 * In the group_imb case we cannot rely on group-wide averages
5526 * to ensure cpu-load equilibrium, look at wider averages. XXX
5528 busiest->load_per_task =
5529 min(busiest->load_per_task, sds->avg_load);
5533 * In the presence of smp nice balancing, certain scenarios can have
5534 * max load less than avg load(as we skip the groups at or below
5535 * its cpu_power, while calculating max_load..)
5537 if (busiest->avg_load <= sds->avg_load ||
5538 local->avg_load >= sds->avg_load) {
5540 return fix_small_imbalance(env, sds);
5543 if (!busiest->group_imb) {
5545 * Don't want to pull so many tasks that a group would go idle.
5546 * Except of course for the group_imb case, since then we might
5547 * have to drop below capacity to reach cpu-load equilibrium.
5549 load_above_capacity =
5550 (busiest->sum_nr_running - busiest->group_capacity);
5552 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5553 load_above_capacity /= busiest->group_power;
5557 * We're trying to get all the cpus to the average_load, so we don't
5558 * want to push ourselves above the average load, nor do we wish to
5559 * reduce the max loaded cpu below the average load. At the same time,
5560 * we also don't want to reduce the group load below the group capacity
5561 * (so that we can implement power-savings policies etc). Thus we look
5562 * for the minimum possible imbalance.
5564 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5566 /* How much load to actually move to equalise the imbalance */
5567 env->imbalance = min(
5568 max_pull * busiest->group_power,
5569 (sds->avg_load - local->avg_load) * local->group_power
5570 ) / SCHED_POWER_SCALE;
5573 * if *imbalance is less than the average load per runnable task
5574 * there is no guarantee that any tasks will be moved so we'll have
5575 * a think about bumping its value to force at least one task to be
5578 if (env->imbalance < busiest->load_per_task)
5579 return fix_small_imbalance(env, sds);
5582 /******* find_busiest_group() helpers end here *********************/
5585 * find_busiest_group - Returns the busiest group within the sched_domain
5586 * if there is an imbalance. If there isn't an imbalance, and
5587 * the user has opted for power-savings, it returns a group whose
5588 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5589 * such a group exists.
5591 * Also calculates the amount of weighted load which should be moved
5592 * to restore balance.
5594 * @env: The load balancing environment.
5596 * Return: - The busiest group if imbalance exists.
5597 * - If no imbalance and user has opted for power-savings balance,
5598 * return the least loaded group whose CPUs can be
5599 * put to idle by rebalancing its tasks onto our group.
5601 static struct sched_group *find_busiest_group(struct lb_env *env)
5603 struct sg_lb_stats *local, *busiest;
5604 struct sd_lb_stats sds;
5606 init_sd_lb_stats(&sds);
5609 * Compute the various statistics relavent for load balancing at
5612 update_sd_lb_stats(env, &sds);
5613 local = &sds.local_stat;
5614 busiest = &sds.busiest_stat;
5616 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5617 check_asym_packing(env, &sds))
5620 /* There is no busy sibling group to pull tasks from */
5621 if (!sds.busiest || busiest->sum_nr_running == 0)
5624 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5627 * If the busiest group is imbalanced the below checks don't
5628 * work because they assume all things are equal, which typically
5629 * isn't true due to cpus_allowed constraints and the like.
5631 if (busiest->group_imb)
5634 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5635 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5636 !busiest->group_has_capacity)
5640 * If the local group is more busy than the selected busiest group
5641 * don't try and pull any tasks.
5643 if (local->avg_load >= busiest->avg_load)
5647 * Don't pull any tasks if this group is already above the domain
5650 if (local->avg_load >= sds.avg_load)
5653 if (env->idle == CPU_IDLE) {
5655 * This cpu is idle. If the busiest group load doesn't
5656 * have more tasks than the number of available cpu's and
5657 * there is no imbalance between this and busiest group
5658 * wrt to idle cpu's, it is balanced.
5660 if ((local->idle_cpus < busiest->idle_cpus) &&
5661 busiest->sum_nr_running <= busiest->group_weight)
5665 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5666 * imbalance_pct to be conservative.
5668 if (100 * busiest->avg_load <=
5669 env->sd->imbalance_pct * local->avg_load)
5674 /* Looks like there is an imbalance. Compute it */
5675 calculate_imbalance(env, &sds);
5684 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5686 static struct rq *find_busiest_queue(struct lb_env *env,
5687 struct sched_group *group)
5689 struct rq *busiest = NULL, *rq;
5690 unsigned long busiest_load = 0, busiest_power = 1;
5693 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5694 unsigned long power = power_of(i);
5695 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5700 capacity = fix_small_capacity(env->sd, group);
5703 wl = weighted_cpuload(i);
5706 * When comparing with imbalance, use weighted_cpuload()
5707 * which is not scaled with the cpu power.
5709 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5713 * For the load comparisons with the other cpu's, consider
5714 * the weighted_cpuload() scaled with the cpu power, so that
5715 * the load can be moved away from the cpu that is potentially
5716 * running at a lower capacity.
5718 * Thus we're looking for max(wl_i / power_i), crosswise
5719 * multiplication to rid ourselves of the division works out
5720 * to: wl_i * power_j > wl_j * power_i; where j is our
5723 if (wl * busiest_power > busiest_load * power) {
5725 busiest_power = power;
5734 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5735 * so long as it is large enough.
5737 #define MAX_PINNED_INTERVAL 512
5739 /* Working cpumask for load_balance and load_balance_newidle. */
5740 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5742 static int need_active_balance(struct lb_env *env)
5744 struct sched_domain *sd = env->sd;
5746 if (env->idle == CPU_NEWLY_IDLE) {
5749 * ASYM_PACKING needs to force migrate tasks from busy but
5750 * higher numbered CPUs in order to pack all tasks in the
5751 * lowest numbered CPUs.
5753 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5757 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5760 static int active_load_balance_cpu_stop(void *data);
5762 static int should_we_balance(struct lb_env *env)
5764 struct sched_group *sg = env->sd->groups;
5765 struct cpumask *sg_cpus, *sg_mask;
5766 int cpu, balance_cpu = -1;
5769 * In the newly idle case, we will allow all the cpu's
5770 * to do the newly idle load balance.
5772 if (env->idle == CPU_NEWLY_IDLE)
5775 sg_cpus = sched_group_cpus(sg);
5776 sg_mask = sched_group_mask(sg);
5777 /* Try to find first idle cpu */
5778 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5779 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5786 if (balance_cpu == -1)
5787 balance_cpu = group_balance_cpu(sg);
5790 * First idle cpu or the first cpu(busiest) in this sched group
5791 * is eligible for doing load balancing at this and above domains.
5793 return balance_cpu == env->dst_cpu;
5797 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5798 * tasks if there is an imbalance.
5800 static int load_balance(int this_cpu, struct rq *this_rq,
5801 struct sched_domain *sd, enum cpu_idle_type idle,
5802 int *continue_balancing)
5804 int ld_moved, cur_ld_moved, active_balance = 0;
5805 struct sched_domain *sd_parent = sd->parent;
5806 struct sched_group *group;
5808 unsigned long flags;
5809 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5811 struct lb_env env = {
5813 .dst_cpu = this_cpu,
5815 .dst_grpmask = sched_group_cpus(sd->groups),
5817 .loop_break = sched_nr_migrate_break,
5822 * For NEWLY_IDLE load_balancing, we don't need to consider
5823 * other cpus in our group
5825 if (idle == CPU_NEWLY_IDLE)
5826 env.dst_grpmask = NULL;
5828 cpumask_copy(cpus, cpu_active_mask);
5830 schedstat_inc(sd, lb_count[idle]);
5833 if (!should_we_balance(&env)) {
5834 *continue_balancing = 0;
5838 group = find_busiest_group(&env);
5840 schedstat_inc(sd, lb_nobusyg[idle]);
5844 busiest = find_busiest_queue(&env, group);
5846 schedstat_inc(sd, lb_nobusyq[idle]);
5850 BUG_ON(busiest == env.dst_rq);
5852 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5855 if (busiest->nr_running > 1) {
5857 * Attempt to move tasks. If find_busiest_group has found
5858 * an imbalance but busiest->nr_running <= 1, the group is
5859 * still unbalanced. ld_moved simply stays zero, so it is
5860 * correctly treated as an imbalance.
5862 env.flags |= LBF_ALL_PINNED;
5863 env.src_cpu = busiest->cpu;
5864 env.src_rq = busiest;
5865 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5868 local_irq_save(flags);
5869 double_rq_lock(env.dst_rq, busiest);
5872 * cur_ld_moved - load moved in current iteration
5873 * ld_moved - cumulative load moved across iterations
5875 cur_ld_moved = move_tasks(&env);
5876 ld_moved += cur_ld_moved;
5877 double_rq_unlock(env.dst_rq, busiest);
5878 local_irq_restore(flags);
5881 * some other cpu did the load balance for us.
5883 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5884 resched_cpu(env.dst_cpu);
5886 if (env.flags & LBF_NEED_BREAK) {
5887 env.flags &= ~LBF_NEED_BREAK;
5892 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5893 * us and move them to an alternate dst_cpu in our sched_group
5894 * where they can run. The upper limit on how many times we
5895 * iterate on same src_cpu is dependent on number of cpus in our
5898 * This changes load balance semantics a bit on who can move
5899 * load to a given_cpu. In addition to the given_cpu itself
5900 * (or a ilb_cpu acting on its behalf where given_cpu is
5901 * nohz-idle), we now have balance_cpu in a position to move
5902 * load to given_cpu. In rare situations, this may cause
5903 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5904 * _independently_ and at _same_ time to move some load to
5905 * given_cpu) causing exceess load to be moved to given_cpu.
5906 * This however should not happen so much in practice and
5907 * moreover subsequent load balance cycles should correct the
5908 * excess load moved.
5910 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5912 /* Prevent to re-select dst_cpu via env's cpus */
5913 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5915 env.dst_rq = cpu_rq(env.new_dst_cpu);
5916 env.dst_cpu = env.new_dst_cpu;
5917 env.flags &= ~LBF_DST_PINNED;
5919 env.loop_break = sched_nr_migrate_break;
5922 * Go back to "more_balance" rather than "redo" since we
5923 * need to continue with same src_cpu.
5929 * We failed to reach balance because of affinity.
5932 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5934 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5935 *group_imbalance = 1;
5936 } else if (*group_imbalance)
5937 *group_imbalance = 0;
5940 /* All tasks on this runqueue were pinned by CPU affinity */
5941 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5942 cpumask_clear_cpu(cpu_of(busiest), cpus);
5943 if (!cpumask_empty(cpus)) {
5945 env.loop_break = sched_nr_migrate_break;
5953 schedstat_inc(sd, lb_failed[idle]);
5955 * Increment the failure counter only on periodic balance.
5956 * We do not want newidle balance, which can be very
5957 * frequent, pollute the failure counter causing
5958 * excessive cache_hot migrations and active balances.
5960 if (idle != CPU_NEWLY_IDLE)
5961 sd->nr_balance_failed++;
5963 if (need_active_balance(&env)) {
5964 raw_spin_lock_irqsave(&busiest->lock, flags);
5966 /* don't kick the active_load_balance_cpu_stop,
5967 * if the curr task on busiest cpu can't be
5970 if (!cpumask_test_cpu(this_cpu,
5971 tsk_cpus_allowed(busiest->curr))) {
5972 raw_spin_unlock_irqrestore(&busiest->lock,
5974 env.flags |= LBF_ALL_PINNED;
5975 goto out_one_pinned;
5979 * ->active_balance synchronizes accesses to
5980 * ->active_balance_work. Once set, it's cleared
5981 * only after active load balance is finished.
5983 if (!busiest->active_balance) {
5984 busiest->active_balance = 1;
5985 busiest->push_cpu = this_cpu;
5988 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5990 if (active_balance) {
5991 stop_one_cpu_nowait(cpu_of(busiest),
5992 active_load_balance_cpu_stop, busiest,
5993 &busiest->active_balance_work);
5997 * We've kicked active balancing, reset the failure
6000 sd->nr_balance_failed = sd->cache_nice_tries+1;
6003 sd->nr_balance_failed = 0;
6005 if (likely(!active_balance)) {
6006 /* We were unbalanced, so reset the balancing interval */
6007 sd->balance_interval = sd->min_interval;
6010 * If we've begun active balancing, start to back off. This
6011 * case may not be covered by the all_pinned logic if there
6012 * is only 1 task on the busy runqueue (because we don't call
6015 if (sd->balance_interval < sd->max_interval)
6016 sd->balance_interval *= 2;
6022 schedstat_inc(sd, lb_balanced[idle]);
6024 sd->nr_balance_failed = 0;
6027 /* tune up the balancing interval */
6028 if (((env.flags & LBF_ALL_PINNED) &&
6029 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6030 (sd->balance_interval < sd->max_interval))
6031 sd->balance_interval *= 2;
6039 * idle_balance is called by schedule() if this_cpu is about to become
6040 * idle. Attempts to pull tasks from other CPUs.
6042 void idle_balance(int this_cpu, struct rq *this_rq)
6044 struct sched_domain *sd;
6045 int pulled_task = 0;
6046 unsigned long next_balance = jiffies + HZ;
6049 this_rq->idle_stamp = rq_clock(this_rq);
6051 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6055 * Drop the rq->lock, but keep IRQ/preempt disabled.
6057 raw_spin_unlock(&this_rq->lock);
6059 update_blocked_averages(this_cpu);
6061 for_each_domain(this_cpu, sd) {
6062 unsigned long interval;
6063 int continue_balancing = 1;
6064 u64 t0, domain_cost;
6066 if (!(sd->flags & SD_LOAD_BALANCE))
6069 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6072 if (sd->flags & SD_BALANCE_NEWIDLE) {
6073 t0 = sched_clock_cpu(this_cpu);
6075 /* If we've pulled tasks over stop searching: */
6076 pulled_task = load_balance(this_cpu, this_rq,
6078 &continue_balancing);
6080 domain_cost = sched_clock_cpu(this_cpu) - t0;
6081 if (domain_cost > sd->max_newidle_lb_cost)
6082 sd->max_newidle_lb_cost = domain_cost;
6084 curr_cost += domain_cost;
6087 interval = msecs_to_jiffies(sd->balance_interval);
6088 if (time_after(next_balance, sd->last_balance + interval))
6089 next_balance = sd->last_balance + interval;
6091 this_rq->idle_stamp = 0;
6097 raw_spin_lock(&this_rq->lock);
6099 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6101 * We are going idle. next_balance may be set based on
6102 * a busy processor. So reset next_balance.
6104 this_rq->next_balance = next_balance;
6107 if (curr_cost > this_rq->max_idle_balance_cost)
6108 this_rq->max_idle_balance_cost = curr_cost;
6112 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6113 * running tasks off the busiest CPU onto idle CPUs. It requires at
6114 * least 1 task to be running on each physical CPU where possible, and
6115 * avoids physical / logical imbalances.
6117 static int active_load_balance_cpu_stop(void *data)
6119 struct rq *busiest_rq = data;
6120 int busiest_cpu = cpu_of(busiest_rq);
6121 int target_cpu = busiest_rq->push_cpu;
6122 struct rq *target_rq = cpu_rq(target_cpu);
6123 struct sched_domain *sd;
6125 raw_spin_lock_irq(&busiest_rq->lock);
6127 /* make sure the requested cpu hasn't gone down in the meantime */
6128 if (unlikely(busiest_cpu != smp_processor_id() ||
6129 !busiest_rq->active_balance))
6132 /* Is there any task to move? */
6133 if (busiest_rq->nr_running <= 1)
6137 * This condition is "impossible", if it occurs
6138 * we need to fix it. Originally reported by
6139 * Bjorn Helgaas on a 128-cpu setup.
6141 BUG_ON(busiest_rq == target_rq);
6143 /* move a task from busiest_rq to target_rq */
6144 double_lock_balance(busiest_rq, target_rq);
6146 /* Search for an sd spanning us and the target CPU. */
6148 for_each_domain(target_cpu, sd) {
6149 if ((sd->flags & SD_LOAD_BALANCE) &&
6150 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6155 struct lb_env env = {
6157 .dst_cpu = target_cpu,
6158 .dst_rq = target_rq,
6159 .src_cpu = busiest_rq->cpu,
6160 .src_rq = busiest_rq,
6164 schedstat_inc(sd, alb_count);
6166 if (move_one_task(&env))
6167 schedstat_inc(sd, alb_pushed);
6169 schedstat_inc(sd, alb_failed);
6172 double_unlock_balance(busiest_rq, target_rq);
6174 busiest_rq->active_balance = 0;
6175 raw_spin_unlock_irq(&busiest_rq->lock);
6179 #ifdef CONFIG_NO_HZ_COMMON
6181 * idle load balancing details
6182 * - When one of the busy CPUs notice that there may be an idle rebalancing
6183 * needed, they will kick the idle load balancer, which then does idle
6184 * load balancing for all the idle CPUs.
6187 cpumask_var_t idle_cpus_mask;
6189 unsigned long next_balance; /* in jiffy units */
6190 } nohz ____cacheline_aligned;
6192 static inline int find_new_ilb(int call_cpu)
6194 int ilb = cpumask_first(nohz.idle_cpus_mask);
6196 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6203 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6204 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6205 * CPU (if there is one).
6207 static void nohz_balancer_kick(int cpu)
6211 nohz.next_balance++;
6213 ilb_cpu = find_new_ilb(cpu);
6215 if (ilb_cpu >= nr_cpu_ids)
6218 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6221 * Use smp_send_reschedule() instead of resched_cpu().
6222 * This way we generate a sched IPI on the target cpu which
6223 * is idle. And the softirq performing nohz idle load balance
6224 * will be run before returning from the IPI.
6226 smp_send_reschedule(ilb_cpu);
6230 static inline void nohz_balance_exit_idle(int cpu)
6232 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6233 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6234 atomic_dec(&nohz.nr_cpus);
6235 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6239 static inline void set_cpu_sd_state_busy(void)
6241 struct sched_domain *sd;
6244 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6246 if (!sd || !sd->nohz_idle)
6250 for (; sd; sd = sd->parent)
6251 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6256 void set_cpu_sd_state_idle(void)
6258 struct sched_domain *sd;
6261 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6263 if (!sd || sd->nohz_idle)
6267 for (; sd; sd = sd->parent)
6268 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6274 * This routine will record that the cpu is going idle with tick stopped.
6275 * This info will be used in performing idle load balancing in the future.
6277 void nohz_balance_enter_idle(int cpu)
6280 * If this cpu is going down, then nothing needs to be done.
6282 if (!cpu_active(cpu))
6285 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6288 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6289 atomic_inc(&nohz.nr_cpus);
6290 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6293 static int sched_ilb_notifier(struct notifier_block *nfb,
6294 unsigned long action, void *hcpu)
6296 switch (action & ~CPU_TASKS_FROZEN) {
6298 nohz_balance_exit_idle(smp_processor_id());
6306 static DEFINE_SPINLOCK(balancing);
6309 * Scale the max load_balance interval with the number of CPUs in the system.
6310 * This trades load-balance latency on larger machines for less cross talk.
6312 void update_max_interval(void)
6314 max_load_balance_interval = HZ*num_online_cpus()/10;
6318 * It checks each scheduling domain to see if it is due to be balanced,
6319 * and initiates a balancing operation if so.
6321 * Balancing parameters are set up in init_sched_domains.
6323 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6325 int continue_balancing = 1;
6326 struct rq *rq = cpu_rq(cpu);
6327 unsigned long interval;
6328 struct sched_domain *sd;
6329 /* Earliest time when we have to do rebalance again */
6330 unsigned long next_balance = jiffies + 60*HZ;
6331 int update_next_balance = 0;
6332 int need_serialize, need_decay = 0;
6335 update_blocked_averages(cpu);
6338 for_each_domain(cpu, sd) {
6340 * Decay the newidle max times here because this is a regular
6341 * visit to all the domains. Decay ~1% per second.
6343 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6344 sd->max_newidle_lb_cost =
6345 (sd->max_newidle_lb_cost * 253) / 256;
6346 sd->next_decay_max_lb_cost = jiffies + HZ;
6349 max_cost += sd->max_newidle_lb_cost;
6351 if (!(sd->flags & SD_LOAD_BALANCE))
6355 * Stop the load balance at this level. There is another
6356 * CPU in our sched group which is doing load balancing more
6359 if (!continue_balancing) {
6365 interval = sd->balance_interval;
6366 if (idle != CPU_IDLE)
6367 interval *= sd->busy_factor;
6369 /* scale ms to jiffies */
6370 interval = msecs_to_jiffies(interval);
6371 interval = clamp(interval, 1UL, max_load_balance_interval);
6373 need_serialize = sd->flags & SD_SERIALIZE;
6375 if (need_serialize) {
6376 if (!spin_trylock(&balancing))
6380 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6381 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6383 * The LBF_DST_PINNED logic could have changed
6384 * env->dst_cpu, so we can't know our idle
6385 * state even if we migrated tasks. Update it.
6387 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6389 sd->last_balance = jiffies;
6392 spin_unlock(&balancing);
6394 if (time_after(next_balance, sd->last_balance + interval)) {
6395 next_balance = sd->last_balance + interval;
6396 update_next_balance = 1;
6401 * Ensure the rq-wide value also decays but keep it at a
6402 * reasonable floor to avoid funnies with rq->avg_idle.
6404 rq->max_idle_balance_cost =
6405 max((u64)sysctl_sched_migration_cost, max_cost);
6410 * next_balance will be updated only when there is a need.
6411 * When the cpu is attached to null domain for ex, it will not be
6414 if (likely(update_next_balance))
6415 rq->next_balance = next_balance;
6418 #ifdef CONFIG_NO_HZ_COMMON
6420 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6421 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6423 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6425 struct rq *this_rq = cpu_rq(this_cpu);
6429 if (idle != CPU_IDLE ||
6430 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6433 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6434 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6438 * If this cpu gets work to do, stop the load balancing
6439 * work being done for other cpus. Next load
6440 * balancing owner will pick it up.
6445 rq = cpu_rq(balance_cpu);
6447 raw_spin_lock_irq(&rq->lock);
6448 update_rq_clock(rq);
6449 update_idle_cpu_load(rq);
6450 raw_spin_unlock_irq(&rq->lock);
6452 rebalance_domains(balance_cpu, CPU_IDLE);
6454 if (time_after(this_rq->next_balance, rq->next_balance))
6455 this_rq->next_balance = rq->next_balance;
6457 nohz.next_balance = this_rq->next_balance;
6459 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6463 * Current heuristic for kicking the idle load balancer in the presence
6464 * of an idle cpu is the system.
6465 * - This rq has more than one task.
6466 * - At any scheduler domain level, this cpu's scheduler group has multiple
6467 * busy cpu's exceeding the group's power.
6468 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6469 * domain span are idle.
6471 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6473 unsigned long now = jiffies;
6474 struct sched_domain *sd;
6476 if (unlikely(idle_cpu(cpu)))
6480 * We may be recently in ticked or tickless idle mode. At the first
6481 * busy tick after returning from idle, we will update the busy stats.
6483 set_cpu_sd_state_busy();
6484 nohz_balance_exit_idle(cpu);
6487 * None are in tickless mode and hence no need for NOHZ idle load
6490 if (likely(!atomic_read(&nohz.nr_cpus)))
6493 if (time_before(now, nohz.next_balance))
6496 if (rq->nr_running >= 2)
6500 for_each_domain(cpu, sd) {
6501 struct sched_group *sg = sd->groups;
6502 struct sched_group_power *sgp = sg->sgp;
6503 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6505 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6506 goto need_kick_unlock;
6508 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6509 && (cpumask_first_and(nohz.idle_cpus_mask,
6510 sched_domain_span(sd)) < cpu))
6511 goto need_kick_unlock;
6513 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6525 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6529 * run_rebalance_domains is triggered when needed from the scheduler tick.
6530 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6532 static void run_rebalance_domains(struct softirq_action *h)
6534 int this_cpu = smp_processor_id();
6535 struct rq *this_rq = cpu_rq(this_cpu);
6536 enum cpu_idle_type idle = this_rq->idle_balance ?
6537 CPU_IDLE : CPU_NOT_IDLE;
6539 rebalance_domains(this_cpu, idle);
6542 * If this cpu has a pending nohz_balance_kick, then do the
6543 * balancing on behalf of the other idle cpus whose ticks are
6546 nohz_idle_balance(this_cpu, idle);
6549 static inline int on_null_domain(int cpu)
6551 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6555 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6557 void trigger_load_balance(struct rq *rq, int cpu)
6559 /* Don't need to rebalance while attached to NULL domain */
6560 if (time_after_eq(jiffies, rq->next_balance) &&
6561 likely(!on_null_domain(cpu)))
6562 raise_softirq(SCHED_SOFTIRQ);
6563 #ifdef CONFIG_NO_HZ_COMMON
6564 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6565 nohz_balancer_kick(cpu);
6569 static void rq_online_fair(struct rq *rq)
6574 static void rq_offline_fair(struct rq *rq)
6578 /* Ensure any throttled groups are reachable by pick_next_task */
6579 unthrottle_offline_cfs_rqs(rq);
6582 #endif /* CONFIG_SMP */
6585 * scheduler tick hitting a task of our scheduling class:
6587 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6589 struct cfs_rq *cfs_rq;
6590 struct sched_entity *se = &curr->se;
6592 for_each_sched_entity(se) {
6593 cfs_rq = cfs_rq_of(se);
6594 entity_tick(cfs_rq, se, queued);
6597 if (numabalancing_enabled)
6598 task_tick_numa(rq, curr);
6600 update_rq_runnable_avg(rq, 1);
6604 * called on fork with the child task as argument from the parent's context
6605 * - child not yet on the tasklist
6606 * - preemption disabled
6608 static void task_fork_fair(struct task_struct *p)
6610 struct cfs_rq *cfs_rq;
6611 struct sched_entity *se = &p->se, *curr;
6612 int this_cpu = smp_processor_id();
6613 struct rq *rq = this_rq();
6614 unsigned long flags;
6616 raw_spin_lock_irqsave(&rq->lock, flags);
6618 update_rq_clock(rq);
6620 cfs_rq = task_cfs_rq(current);
6621 curr = cfs_rq->curr;
6624 * Not only the cpu but also the task_group of the parent might have
6625 * been changed after parent->se.parent,cfs_rq were copied to
6626 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6627 * of child point to valid ones.
6630 __set_task_cpu(p, this_cpu);
6633 update_curr(cfs_rq);
6636 se->vruntime = curr->vruntime;
6637 place_entity(cfs_rq, se, 1);
6639 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6641 * Upon rescheduling, sched_class::put_prev_task() will place
6642 * 'current' within the tree based on its new key value.
6644 swap(curr->vruntime, se->vruntime);
6645 resched_task(rq->curr);
6648 se->vruntime -= cfs_rq->min_vruntime;
6650 raw_spin_unlock_irqrestore(&rq->lock, flags);
6654 * Priority of the task has changed. Check to see if we preempt
6658 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6664 * Reschedule if we are currently running on this runqueue and
6665 * our priority decreased, or if we are not currently running on
6666 * this runqueue and our priority is higher than the current's
6668 if (rq->curr == p) {
6669 if (p->prio > oldprio)
6670 resched_task(rq->curr);
6672 check_preempt_curr(rq, p, 0);
6675 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6677 struct sched_entity *se = &p->se;
6678 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6681 * Ensure the task's vruntime is normalized, so that when its
6682 * switched back to the fair class the enqueue_entity(.flags=0) will
6683 * do the right thing.
6685 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6686 * have normalized the vruntime, if it was !on_rq, then only when
6687 * the task is sleeping will it still have non-normalized vruntime.
6689 if (!se->on_rq && p->state != TASK_RUNNING) {
6691 * Fix up our vruntime so that the current sleep doesn't
6692 * cause 'unlimited' sleep bonus.
6694 place_entity(cfs_rq, se, 0);
6695 se->vruntime -= cfs_rq->min_vruntime;
6700 * Remove our load from contribution when we leave sched_fair
6701 * and ensure we don't carry in an old decay_count if we
6704 if (se->avg.decay_count) {
6705 __synchronize_entity_decay(se);
6706 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6712 * We switched to the sched_fair class.
6714 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6720 * We were most likely switched from sched_rt, so
6721 * kick off the schedule if running, otherwise just see
6722 * if we can still preempt the current task.
6725 resched_task(rq->curr);
6727 check_preempt_curr(rq, p, 0);
6730 /* Account for a task changing its policy or group.
6732 * This routine is mostly called to set cfs_rq->curr field when a task
6733 * migrates between groups/classes.
6735 static void set_curr_task_fair(struct rq *rq)
6737 struct sched_entity *se = &rq->curr->se;
6739 for_each_sched_entity(se) {
6740 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6742 set_next_entity(cfs_rq, se);
6743 /* ensure bandwidth has been allocated on our new cfs_rq */
6744 account_cfs_rq_runtime(cfs_rq, 0);
6748 void init_cfs_rq(struct cfs_rq *cfs_rq)
6750 cfs_rq->tasks_timeline = RB_ROOT;
6751 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6752 #ifndef CONFIG_64BIT
6753 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6756 atomic64_set(&cfs_rq->decay_counter, 1);
6757 atomic_long_set(&cfs_rq->removed_load, 0);
6761 #ifdef CONFIG_FAIR_GROUP_SCHED
6762 static void task_move_group_fair(struct task_struct *p, int on_rq)
6764 struct cfs_rq *cfs_rq;
6766 * If the task was not on the rq at the time of this cgroup movement
6767 * it must have been asleep, sleeping tasks keep their ->vruntime
6768 * absolute on their old rq until wakeup (needed for the fair sleeper
6769 * bonus in place_entity()).
6771 * If it was on the rq, we've just 'preempted' it, which does convert
6772 * ->vruntime to a relative base.
6774 * Make sure both cases convert their relative position when migrating
6775 * to another cgroup's rq. This does somewhat interfere with the
6776 * fair sleeper stuff for the first placement, but who cares.
6779 * When !on_rq, vruntime of the task has usually NOT been normalized.
6780 * But there are some cases where it has already been normalized:
6782 * - Moving a forked child which is waiting for being woken up by
6783 * wake_up_new_task().
6784 * - Moving a task which has been woken up by try_to_wake_up() and
6785 * waiting for actually being woken up by sched_ttwu_pending().
6787 * To prevent boost or penalty in the new cfs_rq caused by delta
6788 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6790 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6794 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6795 set_task_rq(p, task_cpu(p));
6797 cfs_rq = cfs_rq_of(&p->se);
6798 p->se.vruntime += cfs_rq->min_vruntime;
6801 * migrate_task_rq_fair() will have removed our previous
6802 * contribution, but we must synchronize for ongoing future
6805 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6806 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6811 void free_fair_sched_group(struct task_group *tg)
6815 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6817 for_each_possible_cpu(i) {
6819 kfree(tg->cfs_rq[i]);
6828 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6830 struct cfs_rq *cfs_rq;
6831 struct sched_entity *se;
6834 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6837 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6841 tg->shares = NICE_0_LOAD;
6843 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6845 for_each_possible_cpu(i) {
6846 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6847 GFP_KERNEL, cpu_to_node(i));
6851 se = kzalloc_node(sizeof(struct sched_entity),
6852 GFP_KERNEL, cpu_to_node(i));
6856 init_cfs_rq(cfs_rq);
6857 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6868 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6870 struct rq *rq = cpu_rq(cpu);
6871 unsigned long flags;
6874 * Only empty task groups can be destroyed; so we can speculatively
6875 * check on_list without danger of it being re-added.
6877 if (!tg->cfs_rq[cpu]->on_list)
6880 raw_spin_lock_irqsave(&rq->lock, flags);
6881 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6882 raw_spin_unlock_irqrestore(&rq->lock, flags);
6885 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6886 struct sched_entity *se, int cpu,
6887 struct sched_entity *parent)
6889 struct rq *rq = cpu_rq(cpu);
6893 init_cfs_rq_runtime(cfs_rq);
6895 tg->cfs_rq[cpu] = cfs_rq;
6898 /* se could be NULL for root_task_group */
6903 se->cfs_rq = &rq->cfs;
6905 se->cfs_rq = parent->my_q;
6908 update_load_set(&se->load, 0);
6909 se->parent = parent;
6912 static DEFINE_MUTEX(shares_mutex);
6914 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6917 unsigned long flags;
6920 * We can't change the weight of the root cgroup.
6925 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6927 mutex_lock(&shares_mutex);
6928 if (tg->shares == shares)
6931 tg->shares = shares;
6932 for_each_possible_cpu(i) {
6933 struct rq *rq = cpu_rq(i);
6934 struct sched_entity *se;
6937 /* Propagate contribution to hierarchy */
6938 raw_spin_lock_irqsave(&rq->lock, flags);
6940 /* Possible calls to update_curr() need rq clock */
6941 update_rq_clock(rq);
6942 for_each_sched_entity(se)
6943 update_cfs_shares(group_cfs_rq(se));
6944 raw_spin_unlock_irqrestore(&rq->lock, flags);
6948 mutex_unlock(&shares_mutex);
6951 #else /* CONFIG_FAIR_GROUP_SCHED */
6953 void free_fair_sched_group(struct task_group *tg) { }
6955 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6960 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6962 #endif /* CONFIG_FAIR_GROUP_SCHED */
6965 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6967 struct sched_entity *se = &task->se;
6968 unsigned int rr_interval = 0;
6971 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6974 if (rq->cfs.load.weight)
6975 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6981 * All the scheduling class methods:
6983 const struct sched_class fair_sched_class = {
6984 .next = &idle_sched_class,
6985 .enqueue_task = enqueue_task_fair,
6986 .dequeue_task = dequeue_task_fair,
6987 .yield_task = yield_task_fair,
6988 .yield_to_task = yield_to_task_fair,
6990 .check_preempt_curr = check_preempt_wakeup,
6992 .pick_next_task = pick_next_task_fair,
6993 .put_prev_task = put_prev_task_fair,
6996 .select_task_rq = select_task_rq_fair,
6997 .migrate_task_rq = migrate_task_rq_fair,
6999 .rq_online = rq_online_fair,
7000 .rq_offline = rq_offline_fair,
7002 .task_waking = task_waking_fair,
7005 .set_curr_task = set_curr_task_fair,
7006 .task_tick = task_tick_fair,
7007 .task_fork = task_fork_fair,
7009 .prio_changed = prio_changed_fair,
7010 .switched_from = switched_from_fair,
7011 .switched_to = switched_to_fair,
7013 .get_rr_interval = get_rr_interval_fair,
7015 #ifdef CONFIG_FAIR_GROUP_SCHED
7016 .task_move_group = task_move_group_fair,
7020 #ifdef CONFIG_SCHED_DEBUG
7021 void print_cfs_stats(struct seq_file *m, int cpu)
7023 struct cfs_rq *cfs_rq;
7026 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7027 print_cfs_rq(m, cpu, cfs_rq);
7032 __init void init_sched_fair_class(void)
7035 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7037 #ifdef CONFIG_NO_HZ_COMMON
7038 nohz.next_balance = jiffies;
7039 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7040 cpu_notifier(sched_ilb_notifier, 0);