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 */
897 struct list_head task_list;
900 atomic_long_t faults[0];
903 pid_t task_numa_group_id(struct task_struct *p)
905 return p->numa_group ? p->numa_group->gid : 0;
908 static inline int task_faults_idx(int nid, int priv)
910 return 2 * nid + priv;
913 static inline unsigned long task_faults(struct task_struct *p, int nid)
918 return p->numa_faults[task_faults_idx(nid, 0)] +
919 p->numa_faults[task_faults_idx(nid, 1)];
922 static unsigned long weighted_cpuload(const int cpu);
923 static unsigned long source_load(int cpu, int type);
924 static unsigned long target_load(int cpu, int type);
925 static unsigned long power_of(int cpu);
926 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
928 /* Cached statistics for all CPUs within a node */
930 unsigned long nr_running;
933 /* Total compute capacity of CPUs on a node */
936 /* Approximate capacity in terms of runnable tasks on a node */
937 unsigned long capacity;
942 * XXX borrowed from update_sg_lb_stats
944 static void update_numa_stats(struct numa_stats *ns, int nid)
948 memset(ns, 0, sizeof(*ns));
949 for_each_cpu(cpu, cpumask_of_node(nid)) {
950 struct rq *rq = cpu_rq(cpu);
952 ns->nr_running += rq->nr_running;
953 ns->load += weighted_cpuload(cpu);
954 ns->power += power_of(cpu);
957 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
958 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
959 ns->has_capacity = (ns->nr_running < ns->capacity);
962 struct task_numa_env {
963 struct task_struct *p;
965 int src_cpu, src_nid;
966 int dst_cpu, dst_nid;
968 struct numa_stats src_stats, dst_stats;
970 int imbalance_pct, idx;
972 struct task_struct *best_task;
977 static void task_numa_assign(struct task_numa_env *env,
978 struct task_struct *p, long imp)
981 put_task_struct(env->best_task);
987 env->best_cpu = env->dst_cpu;
991 * This checks if the overall compute and NUMA accesses of the system would
992 * be improved if the source tasks was migrated to the target dst_cpu taking
993 * into account that it might be best if task running on the dst_cpu should
994 * be exchanged with the source task
996 static void task_numa_compare(struct task_numa_env *env, long imp)
998 struct rq *src_rq = cpu_rq(env->src_cpu);
999 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1000 struct task_struct *cur;
1001 long dst_load, src_load;
1005 cur = ACCESS_ONCE(dst_rq->curr);
1006 if (cur->pid == 0) /* idle */
1010 * "imp" is the fault differential for the source task between the
1011 * source and destination node. Calculate the total differential for
1012 * the source task and potential destination task. The more negative
1013 * the value is, the more rmeote accesses that would be expected to
1014 * be incurred if the tasks were swapped.
1017 /* Skip this swap candidate if cannot move to the source cpu */
1018 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1021 imp += task_faults(cur, env->src_nid) -
1022 task_faults(cur, env->dst_nid);
1025 if (imp < env->best_imp)
1029 /* Is there capacity at our destination? */
1030 if (env->src_stats.has_capacity &&
1031 !env->dst_stats.has_capacity)
1037 /* Balance doesn't matter much if we're running a task per cpu */
1038 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1042 * In the overloaded case, try and keep the load balanced.
1045 dst_load = env->dst_stats.load;
1046 src_load = env->src_stats.load;
1048 /* XXX missing power terms */
1049 load = task_h_load(env->p);
1054 load = task_h_load(cur);
1059 /* make src_load the smaller */
1060 if (dst_load < src_load)
1061 swap(dst_load, src_load);
1063 if (src_load * env->imbalance_pct < dst_load * 100)
1067 task_numa_assign(env, cur, imp);
1072 static void task_numa_find_cpu(struct task_numa_env *env, long imp)
1076 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1077 /* Skip this CPU if the source task cannot migrate */
1078 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1082 task_numa_compare(env, imp);
1086 static int task_numa_migrate(struct task_struct *p)
1088 struct task_numa_env env = {
1091 .src_cpu = task_cpu(p),
1092 .src_nid = cpu_to_node(task_cpu(p)),
1094 .imbalance_pct = 112,
1100 struct sched_domain *sd;
1101 unsigned long faults;
1106 * Pick the lowest SD_NUMA domain, as that would have the smallest
1107 * imbalance and would be the first to start moving tasks about.
1109 * And we want to avoid any moving of tasks about, as that would create
1110 * random movement of tasks -- counter the numa conditions we're trying
1114 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1115 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1118 faults = task_faults(p, env.src_nid);
1119 update_numa_stats(&env.src_stats, env.src_nid);
1120 env.dst_nid = p->numa_preferred_nid;
1121 imp = task_faults(env.p, env.dst_nid) - faults;
1122 update_numa_stats(&env.dst_stats, env.dst_nid);
1124 /* If the preferred nid has capacity, try to use it. */
1125 if (env.dst_stats.has_capacity)
1126 task_numa_find_cpu(&env, imp);
1128 /* No space available on the preferred nid. Look elsewhere. */
1129 if (env.best_cpu == -1) {
1130 for_each_online_node(nid) {
1131 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1134 /* Only consider nodes that recorded more faults */
1135 imp = task_faults(env.p, nid) - faults;
1140 update_numa_stats(&env.dst_stats, env.dst_nid);
1141 task_numa_find_cpu(&env, imp);
1145 /* No better CPU than the current one was found. */
1146 if (env.best_cpu == -1)
1149 if (env.best_task == NULL) {
1150 int ret = migrate_task_to(p, env.best_cpu);
1154 ret = migrate_swap(p, env.best_task);
1155 put_task_struct(env.best_task);
1159 /* Attempt to migrate a task to a CPU on the preferred node. */
1160 static void numa_migrate_preferred(struct task_struct *p)
1162 /* Success if task is already running on preferred CPU */
1163 p->numa_migrate_retry = 0;
1164 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1166 * If migration is temporarily disabled due to a task migration
1167 * then re-enable it now as the task is running on its
1168 * preferred node and memory should migrate locally
1170 if (!p->numa_migrate_seq)
1171 p->numa_migrate_seq++;
1175 /* This task has no NUMA fault statistics yet */
1176 if (unlikely(p->numa_preferred_nid == -1))
1179 /* Otherwise, try migrate to a CPU on the preferred node */
1180 if (task_numa_migrate(p) != 0)
1181 p->numa_migrate_retry = jiffies + HZ*5;
1184 static void task_numa_placement(struct task_struct *p)
1186 int seq, nid, max_nid = -1;
1187 unsigned long max_faults = 0;
1189 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1190 if (p->numa_scan_seq == seq)
1192 p->numa_scan_seq = seq;
1193 p->numa_migrate_seq++;
1194 p->numa_scan_period_max = task_scan_max(p);
1196 /* Find the node with the highest number of faults */
1197 for_each_online_node(nid) {
1198 unsigned long faults = 0;
1201 for (priv = 0; priv < 2; priv++) {
1204 i = task_faults_idx(nid, priv);
1205 diff = -p->numa_faults[i];
1207 /* Decay existing window, copy faults since last scan */
1208 p->numa_faults[i] >>= 1;
1209 p->numa_faults[i] += p->numa_faults_buffer[i];
1210 p->numa_faults_buffer[i] = 0;
1212 faults += p->numa_faults[i];
1213 diff += p->numa_faults[i];
1214 if (p->numa_group) {
1215 /* safe because we can only change our own group */
1216 atomic_long_add(diff, &p->numa_group->faults[i]);
1220 if (faults > max_faults) {
1221 max_faults = faults;
1226 /* Preferred node as the node with the most faults */
1227 if (max_faults && max_nid != p->numa_preferred_nid) {
1228 /* Update the preferred nid and migrate task if possible */
1229 p->numa_preferred_nid = max_nid;
1230 p->numa_migrate_seq = 1;
1231 numa_migrate_preferred(p);
1235 static inline int get_numa_group(struct numa_group *grp)
1237 return atomic_inc_not_zero(&grp->refcount);
1240 static inline void put_numa_group(struct numa_group *grp)
1242 if (atomic_dec_and_test(&grp->refcount))
1243 kfree_rcu(grp, rcu);
1246 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1252 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1255 static void task_numa_group(struct task_struct *p, int cpupid)
1257 struct numa_group *grp, *my_grp;
1258 struct task_struct *tsk;
1260 int cpu = cpupid_to_cpu(cpupid);
1263 if (unlikely(!p->numa_group)) {
1264 unsigned int size = sizeof(struct numa_group) +
1265 2*nr_node_ids*sizeof(atomic_long_t);
1267 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1271 atomic_set(&grp->refcount, 1);
1272 spin_lock_init(&grp->lock);
1273 INIT_LIST_HEAD(&grp->task_list);
1276 for (i = 0; i < 2*nr_node_ids; i++)
1277 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1279 list_add(&p->numa_entry, &grp->task_list);
1281 rcu_assign_pointer(p->numa_group, grp);
1285 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1287 if (!cpupid_match_pid(tsk, cpupid))
1290 grp = rcu_dereference(tsk->numa_group);
1294 my_grp = p->numa_group;
1299 * Only join the other group if its bigger; if we're the bigger group,
1300 * the other task will join us.
1302 if (my_grp->nr_tasks > grp->nr_tasks)
1306 * Tie-break on the grp address.
1308 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1311 if (!get_numa_group(grp))
1322 for (i = 0; i < 2*nr_node_ids; i++) {
1323 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1324 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1327 double_lock(&my_grp->lock, &grp->lock);
1329 list_move(&p->numa_entry, &grp->task_list);
1333 spin_unlock(&my_grp->lock);
1334 spin_unlock(&grp->lock);
1336 rcu_assign_pointer(p->numa_group, grp);
1338 put_numa_group(my_grp);
1341 void task_numa_free(struct task_struct *p)
1343 struct numa_group *grp = p->numa_group;
1347 for (i = 0; i < 2*nr_node_ids; i++)
1348 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1350 spin_lock(&grp->lock);
1351 list_del(&p->numa_entry);
1353 spin_unlock(&grp->lock);
1354 rcu_assign_pointer(p->numa_group, NULL);
1355 put_numa_group(grp);
1358 kfree(p->numa_faults);
1362 * Got a PROT_NONE fault for a page on @node.
1364 void task_numa_fault(int last_cpupid, int node, int pages, bool migrated)
1366 struct task_struct *p = current;
1369 if (!numabalancing_enabled)
1372 /* for example, ksmd faulting in a user's mm */
1376 /* Allocate buffer to track faults on a per-node basis */
1377 if (unlikely(!p->numa_faults)) {
1378 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1380 /* numa_faults and numa_faults_buffer share the allocation */
1381 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1382 if (!p->numa_faults)
1385 BUG_ON(p->numa_faults_buffer);
1386 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1390 * First accesses are treated as private, otherwise consider accesses
1391 * to be private if the accessing pid has not changed
1393 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1396 priv = cpupid_match_pid(p, last_cpupid);
1398 task_numa_group(p, last_cpupid);
1402 * If pages are properly placed (did not migrate) then scan slower.
1403 * This is reset periodically in case of phase changes
1406 /* Initialise if necessary */
1407 if (!p->numa_scan_period_max)
1408 p->numa_scan_period_max = task_scan_max(p);
1410 p->numa_scan_period = min(p->numa_scan_period_max,
1411 p->numa_scan_period + 10);
1414 task_numa_placement(p);
1416 /* Retry task to preferred node migration if it previously failed */
1417 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1418 numa_migrate_preferred(p);
1420 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1423 static void reset_ptenuma_scan(struct task_struct *p)
1425 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1426 p->mm->numa_scan_offset = 0;
1430 * The expensive part of numa migration is done from task_work context.
1431 * Triggered from task_tick_numa().
1433 void task_numa_work(struct callback_head *work)
1435 unsigned long migrate, next_scan, now = jiffies;
1436 struct task_struct *p = current;
1437 struct mm_struct *mm = p->mm;
1438 struct vm_area_struct *vma;
1439 unsigned long start, end;
1440 unsigned long nr_pte_updates = 0;
1443 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1445 work->next = work; /* protect against double add */
1447 * Who cares about NUMA placement when they're dying.
1449 * NOTE: make sure not to dereference p->mm before this check,
1450 * exit_task_work() happens _after_ exit_mm() so we could be called
1451 * without p->mm even though we still had it when we enqueued this
1454 if (p->flags & PF_EXITING)
1457 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1458 mm->numa_next_scan = now +
1459 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1460 mm->numa_next_reset = now +
1461 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1465 * Reset the scan period if enough time has gone by. Objective is that
1466 * scanning will be reduced if pages are properly placed. As tasks
1467 * can enter different phases this needs to be re-examined. Lacking
1468 * proper tracking of reference behaviour, this blunt hammer is used.
1470 migrate = mm->numa_next_reset;
1471 if (time_after(now, migrate)) {
1472 p->numa_scan_period = task_scan_min(p);
1473 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1474 xchg(&mm->numa_next_reset, next_scan);
1478 * Enforce maximal scan/migration frequency..
1480 migrate = mm->numa_next_scan;
1481 if (time_before(now, migrate))
1484 if (p->numa_scan_period == 0) {
1485 p->numa_scan_period_max = task_scan_max(p);
1486 p->numa_scan_period = task_scan_min(p);
1489 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1490 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1494 * Delay this task enough that another task of this mm will likely win
1495 * the next time around.
1497 p->node_stamp += 2 * TICK_NSEC;
1499 start = mm->numa_scan_offset;
1500 pages = sysctl_numa_balancing_scan_size;
1501 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1505 down_read(&mm->mmap_sem);
1506 vma = find_vma(mm, start);
1508 reset_ptenuma_scan(p);
1512 for (; vma; vma = vma->vm_next) {
1513 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1517 * Shared library pages mapped by multiple processes are not
1518 * migrated as it is expected they are cache replicated. Avoid
1519 * hinting faults in read-only file-backed mappings or the vdso
1520 * as migrating the pages will be of marginal benefit.
1523 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1527 start = max(start, vma->vm_start);
1528 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1529 end = min(end, vma->vm_end);
1530 nr_pte_updates += change_prot_numa(vma, start, end);
1533 * Scan sysctl_numa_balancing_scan_size but ensure that
1534 * at least one PTE is updated so that unused virtual
1535 * address space is quickly skipped.
1538 pages -= (end - start) >> PAGE_SHIFT;
1543 } while (end != vma->vm_end);
1548 * If the whole process was scanned without updates then no NUMA
1549 * hinting faults are being recorded and scan rate should be lower.
1551 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1552 p->numa_scan_period = min(p->numa_scan_period_max,
1553 p->numa_scan_period << 1);
1555 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1556 mm->numa_next_scan = next_scan;
1560 * It is possible to reach the end of the VMA list but the last few
1561 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1562 * would find the !migratable VMA on the next scan but not reset the
1563 * scanner to the start so check it now.
1566 mm->numa_scan_offset = start;
1568 reset_ptenuma_scan(p);
1569 up_read(&mm->mmap_sem);
1573 * Drive the periodic memory faults..
1575 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1577 struct callback_head *work = &curr->numa_work;
1581 * We don't care about NUMA placement if we don't have memory.
1583 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1587 * Using runtime rather than walltime has the dual advantage that
1588 * we (mostly) drive the selection from busy threads and that the
1589 * task needs to have done some actual work before we bother with
1592 now = curr->se.sum_exec_runtime;
1593 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1595 if (now - curr->node_stamp > period) {
1596 if (!curr->node_stamp)
1597 curr->numa_scan_period = task_scan_min(curr);
1598 curr->node_stamp += period;
1600 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1601 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1602 task_work_add(curr, work, true);
1607 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1610 #endif /* CONFIG_NUMA_BALANCING */
1613 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1615 update_load_add(&cfs_rq->load, se->load.weight);
1616 if (!parent_entity(se))
1617 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1619 if (entity_is_task(se))
1620 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1622 cfs_rq->nr_running++;
1626 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1628 update_load_sub(&cfs_rq->load, se->load.weight);
1629 if (!parent_entity(se))
1630 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1631 if (entity_is_task(se))
1632 list_del_init(&se->group_node);
1633 cfs_rq->nr_running--;
1636 #ifdef CONFIG_FAIR_GROUP_SCHED
1638 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1643 * Use this CPU's actual weight instead of the last load_contribution
1644 * to gain a more accurate current total weight. See
1645 * update_cfs_rq_load_contribution().
1647 tg_weight = atomic_long_read(&tg->load_avg);
1648 tg_weight -= cfs_rq->tg_load_contrib;
1649 tg_weight += cfs_rq->load.weight;
1654 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1656 long tg_weight, load, shares;
1658 tg_weight = calc_tg_weight(tg, cfs_rq);
1659 load = cfs_rq->load.weight;
1661 shares = (tg->shares * load);
1663 shares /= tg_weight;
1665 if (shares < MIN_SHARES)
1666 shares = MIN_SHARES;
1667 if (shares > tg->shares)
1668 shares = tg->shares;
1672 # else /* CONFIG_SMP */
1673 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1677 # endif /* CONFIG_SMP */
1678 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1679 unsigned long weight)
1682 /* commit outstanding execution time */
1683 if (cfs_rq->curr == se)
1684 update_curr(cfs_rq);
1685 account_entity_dequeue(cfs_rq, se);
1688 update_load_set(&se->load, weight);
1691 account_entity_enqueue(cfs_rq, se);
1694 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1696 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1698 struct task_group *tg;
1699 struct sched_entity *se;
1703 se = tg->se[cpu_of(rq_of(cfs_rq))];
1704 if (!se || throttled_hierarchy(cfs_rq))
1707 if (likely(se->load.weight == tg->shares))
1710 shares = calc_cfs_shares(cfs_rq, tg);
1712 reweight_entity(cfs_rq_of(se), se, shares);
1714 #else /* CONFIG_FAIR_GROUP_SCHED */
1715 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1718 #endif /* CONFIG_FAIR_GROUP_SCHED */
1722 * We choose a half-life close to 1 scheduling period.
1723 * Note: The tables below are dependent on this value.
1725 #define LOAD_AVG_PERIOD 32
1726 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1727 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1729 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1730 static const u32 runnable_avg_yN_inv[] = {
1731 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1732 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1733 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1734 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1735 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1736 0x85aac367, 0x82cd8698,
1740 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1741 * over-estimates when re-combining.
1743 static const u32 runnable_avg_yN_sum[] = {
1744 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1745 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1746 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1751 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1753 static __always_inline u64 decay_load(u64 val, u64 n)
1755 unsigned int local_n;
1759 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1762 /* after bounds checking we can collapse to 32-bit */
1766 * As y^PERIOD = 1/2, we can combine
1767 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1768 * With a look-up table which covers k^n (n<PERIOD)
1770 * To achieve constant time decay_load.
1772 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1773 val >>= local_n / LOAD_AVG_PERIOD;
1774 local_n %= LOAD_AVG_PERIOD;
1777 val *= runnable_avg_yN_inv[local_n];
1778 /* We don't use SRR here since we always want to round down. */
1783 * For updates fully spanning n periods, the contribution to runnable
1784 * average will be: \Sum 1024*y^n
1786 * We can compute this reasonably efficiently by combining:
1787 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1789 static u32 __compute_runnable_contrib(u64 n)
1793 if (likely(n <= LOAD_AVG_PERIOD))
1794 return runnable_avg_yN_sum[n];
1795 else if (unlikely(n >= LOAD_AVG_MAX_N))
1796 return LOAD_AVG_MAX;
1798 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1800 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1801 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1803 n -= LOAD_AVG_PERIOD;
1804 } while (n > LOAD_AVG_PERIOD);
1806 contrib = decay_load(contrib, n);
1807 return contrib + runnable_avg_yN_sum[n];
1811 * We can represent the historical contribution to runnable average as the
1812 * coefficients of a geometric series. To do this we sub-divide our runnable
1813 * history into segments of approximately 1ms (1024us); label the segment that
1814 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1816 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1818 * (now) (~1ms ago) (~2ms ago)
1820 * Let u_i denote the fraction of p_i that the entity was runnable.
1822 * We then designate the fractions u_i as our co-efficients, yielding the
1823 * following representation of historical load:
1824 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1826 * We choose y based on the with of a reasonably scheduling period, fixing:
1829 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1830 * approximately half as much as the contribution to load within the last ms
1833 * When a period "rolls over" and we have new u_0`, multiplying the previous
1834 * sum again by y is sufficient to update:
1835 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1836 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1838 static __always_inline int __update_entity_runnable_avg(u64 now,
1839 struct sched_avg *sa,
1843 u32 runnable_contrib;
1844 int delta_w, decayed = 0;
1846 delta = now - sa->last_runnable_update;
1848 * This should only happen when time goes backwards, which it
1849 * unfortunately does during sched clock init when we swap over to TSC.
1851 if ((s64)delta < 0) {
1852 sa->last_runnable_update = now;
1857 * Use 1024ns as the unit of measurement since it's a reasonable
1858 * approximation of 1us and fast to compute.
1863 sa->last_runnable_update = now;
1865 /* delta_w is the amount already accumulated against our next period */
1866 delta_w = sa->runnable_avg_period % 1024;
1867 if (delta + delta_w >= 1024) {
1868 /* period roll-over */
1872 * Now that we know we're crossing a period boundary, figure
1873 * out how much from delta we need to complete the current
1874 * period and accrue it.
1876 delta_w = 1024 - delta_w;
1878 sa->runnable_avg_sum += delta_w;
1879 sa->runnable_avg_period += delta_w;
1883 /* Figure out how many additional periods this update spans */
1884 periods = delta / 1024;
1887 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1889 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1892 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1893 runnable_contrib = __compute_runnable_contrib(periods);
1895 sa->runnable_avg_sum += runnable_contrib;
1896 sa->runnable_avg_period += runnable_contrib;
1899 /* Remainder of delta accrued against u_0` */
1901 sa->runnable_avg_sum += delta;
1902 sa->runnable_avg_period += delta;
1907 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1908 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1910 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1911 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1913 decays -= se->avg.decay_count;
1917 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1918 se->avg.decay_count = 0;
1923 #ifdef CONFIG_FAIR_GROUP_SCHED
1924 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1927 struct task_group *tg = cfs_rq->tg;
1930 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1931 tg_contrib -= cfs_rq->tg_load_contrib;
1933 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1934 atomic_long_add(tg_contrib, &tg->load_avg);
1935 cfs_rq->tg_load_contrib += tg_contrib;
1940 * Aggregate cfs_rq runnable averages into an equivalent task_group
1941 * representation for computing load contributions.
1943 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1944 struct cfs_rq *cfs_rq)
1946 struct task_group *tg = cfs_rq->tg;
1949 /* The fraction of a cpu used by this cfs_rq */
1950 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1951 sa->runnable_avg_period + 1);
1952 contrib -= cfs_rq->tg_runnable_contrib;
1954 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1955 atomic_add(contrib, &tg->runnable_avg);
1956 cfs_rq->tg_runnable_contrib += contrib;
1960 static inline void __update_group_entity_contrib(struct sched_entity *se)
1962 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1963 struct task_group *tg = cfs_rq->tg;
1968 contrib = cfs_rq->tg_load_contrib * tg->shares;
1969 se->avg.load_avg_contrib = div_u64(contrib,
1970 atomic_long_read(&tg->load_avg) + 1);
1973 * For group entities we need to compute a correction term in the case
1974 * that they are consuming <1 cpu so that we would contribute the same
1975 * load as a task of equal weight.
1977 * Explicitly co-ordinating this measurement would be expensive, but
1978 * fortunately the sum of each cpus contribution forms a usable
1979 * lower-bound on the true value.
1981 * Consider the aggregate of 2 contributions. Either they are disjoint
1982 * (and the sum represents true value) or they are disjoint and we are
1983 * understating by the aggregate of their overlap.
1985 * Extending this to N cpus, for a given overlap, the maximum amount we
1986 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1987 * cpus that overlap for this interval and w_i is the interval width.
1989 * On a small machine; the first term is well-bounded which bounds the
1990 * total error since w_i is a subset of the period. Whereas on a
1991 * larger machine, while this first term can be larger, if w_i is the
1992 * of consequential size guaranteed to see n_i*w_i quickly converge to
1993 * our upper bound of 1-cpu.
1995 runnable_avg = atomic_read(&tg->runnable_avg);
1996 if (runnable_avg < NICE_0_LOAD) {
1997 se->avg.load_avg_contrib *= runnable_avg;
1998 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2002 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2003 int force_update) {}
2004 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2005 struct cfs_rq *cfs_rq) {}
2006 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2009 static inline void __update_task_entity_contrib(struct sched_entity *se)
2013 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2014 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2015 contrib /= (se->avg.runnable_avg_period + 1);
2016 se->avg.load_avg_contrib = scale_load(contrib);
2019 /* Compute the current contribution to load_avg by se, return any delta */
2020 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2022 long old_contrib = se->avg.load_avg_contrib;
2024 if (entity_is_task(se)) {
2025 __update_task_entity_contrib(se);
2027 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2028 __update_group_entity_contrib(se);
2031 return se->avg.load_avg_contrib - old_contrib;
2034 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2037 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2038 cfs_rq->blocked_load_avg -= load_contrib;
2040 cfs_rq->blocked_load_avg = 0;
2043 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2045 /* Update a sched_entity's runnable average */
2046 static inline void update_entity_load_avg(struct sched_entity *se,
2049 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2054 * For a group entity we need to use their owned cfs_rq_clock_task() in
2055 * case they are the parent of a throttled hierarchy.
2057 if (entity_is_task(se))
2058 now = cfs_rq_clock_task(cfs_rq);
2060 now = cfs_rq_clock_task(group_cfs_rq(se));
2062 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2065 contrib_delta = __update_entity_load_avg_contrib(se);
2071 cfs_rq->runnable_load_avg += contrib_delta;
2073 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2077 * Decay the load contributed by all blocked children and account this so that
2078 * their contribution may appropriately discounted when they wake up.
2080 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2082 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2085 decays = now - cfs_rq->last_decay;
2086 if (!decays && !force_update)
2089 if (atomic_long_read(&cfs_rq->removed_load)) {
2090 unsigned long removed_load;
2091 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2092 subtract_blocked_load_contrib(cfs_rq, removed_load);
2096 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2098 atomic64_add(decays, &cfs_rq->decay_counter);
2099 cfs_rq->last_decay = now;
2102 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2105 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2107 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2108 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2111 /* Add the load generated by se into cfs_rq's child load-average */
2112 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2113 struct sched_entity *se,
2117 * We track migrations using entity decay_count <= 0, on a wake-up
2118 * migration we use a negative decay count to track the remote decays
2119 * accumulated while sleeping.
2121 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2122 * are seen by enqueue_entity_load_avg() as a migration with an already
2123 * constructed load_avg_contrib.
2125 if (unlikely(se->avg.decay_count <= 0)) {
2126 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2127 if (se->avg.decay_count) {
2129 * In a wake-up migration we have to approximate the
2130 * time sleeping. This is because we can't synchronize
2131 * clock_task between the two cpus, and it is not
2132 * guaranteed to be read-safe. Instead, we can
2133 * approximate this using our carried decays, which are
2134 * explicitly atomically readable.
2136 se->avg.last_runnable_update -= (-se->avg.decay_count)
2138 update_entity_load_avg(se, 0);
2139 /* Indicate that we're now synchronized and on-rq */
2140 se->avg.decay_count = 0;
2145 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2146 * would have made count negative); we must be careful to avoid
2147 * double-accounting blocked time after synchronizing decays.
2149 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2153 /* migrated tasks did not contribute to our blocked load */
2155 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2156 update_entity_load_avg(se, 0);
2159 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2160 /* we force update consideration on load-balancer moves */
2161 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2165 * Remove se's load from this cfs_rq child load-average, if the entity is
2166 * transitioning to a blocked state we track its projected decay using
2169 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2170 struct sched_entity *se,
2173 update_entity_load_avg(se, 1);
2174 /* we force update consideration on load-balancer moves */
2175 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2177 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2179 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2180 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2181 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2185 * Update the rq's load with the elapsed running time before entering
2186 * idle. if the last scheduled task is not a CFS task, idle_enter will
2187 * be the only way to update the runnable statistic.
2189 void idle_enter_fair(struct rq *this_rq)
2191 update_rq_runnable_avg(this_rq, 1);
2195 * Update the rq's load with the elapsed idle time before a task is
2196 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2197 * be the only way to update the runnable statistic.
2199 void idle_exit_fair(struct rq *this_rq)
2201 update_rq_runnable_avg(this_rq, 0);
2205 static inline void update_entity_load_avg(struct sched_entity *se,
2206 int update_cfs_rq) {}
2207 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2208 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2209 struct sched_entity *se,
2211 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2212 struct sched_entity *se,
2214 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2215 int force_update) {}
2218 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2220 #ifdef CONFIG_SCHEDSTATS
2221 struct task_struct *tsk = NULL;
2223 if (entity_is_task(se))
2226 if (se->statistics.sleep_start) {
2227 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2232 if (unlikely(delta > se->statistics.sleep_max))
2233 se->statistics.sleep_max = delta;
2235 se->statistics.sleep_start = 0;
2236 se->statistics.sum_sleep_runtime += delta;
2239 account_scheduler_latency(tsk, delta >> 10, 1);
2240 trace_sched_stat_sleep(tsk, delta);
2243 if (se->statistics.block_start) {
2244 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2249 if (unlikely(delta > se->statistics.block_max))
2250 se->statistics.block_max = delta;
2252 se->statistics.block_start = 0;
2253 se->statistics.sum_sleep_runtime += delta;
2256 if (tsk->in_iowait) {
2257 se->statistics.iowait_sum += delta;
2258 se->statistics.iowait_count++;
2259 trace_sched_stat_iowait(tsk, delta);
2262 trace_sched_stat_blocked(tsk, delta);
2265 * Blocking time is in units of nanosecs, so shift by
2266 * 20 to get a milliseconds-range estimation of the
2267 * amount of time that the task spent sleeping:
2269 if (unlikely(prof_on == SLEEP_PROFILING)) {
2270 profile_hits(SLEEP_PROFILING,
2271 (void *)get_wchan(tsk),
2274 account_scheduler_latency(tsk, delta >> 10, 0);
2280 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2282 #ifdef CONFIG_SCHED_DEBUG
2283 s64 d = se->vruntime - cfs_rq->min_vruntime;
2288 if (d > 3*sysctl_sched_latency)
2289 schedstat_inc(cfs_rq, nr_spread_over);
2294 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2296 u64 vruntime = cfs_rq->min_vruntime;
2299 * The 'current' period is already promised to the current tasks,
2300 * however the extra weight of the new task will slow them down a
2301 * little, place the new task so that it fits in the slot that
2302 * stays open at the end.
2304 if (initial && sched_feat(START_DEBIT))
2305 vruntime += sched_vslice(cfs_rq, se);
2307 /* sleeps up to a single latency don't count. */
2309 unsigned long thresh = sysctl_sched_latency;
2312 * Halve their sleep time's effect, to allow
2313 * for a gentler effect of sleepers:
2315 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2321 /* ensure we never gain time by being placed backwards. */
2322 se->vruntime = max_vruntime(se->vruntime, vruntime);
2325 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2328 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2331 * Update the normalized vruntime before updating min_vruntime
2332 * through calling update_curr().
2334 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2335 se->vruntime += cfs_rq->min_vruntime;
2338 * Update run-time statistics of the 'current'.
2340 update_curr(cfs_rq);
2341 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2342 account_entity_enqueue(cfs_rq, se);
2343 update_cfs_shares(cfs_rq);
2345 if (flags & ENQUEUE_WAKEUP) {
2346 place_entity(cfs_rq, se, 0);
2347 enqueue_sleeper(cfs_rq, se);
2350 update_stats_enqueue(cfs_rq, se);
2351 check_spread(cfs_rq, se);
2352 if (se != cfs_rq->curr)
2353 __enqueue_entity(cfs_rq, se);
2356 if (cfs_rq->nr_running == 1) {
2357 list_add_leaf_cfs_rq(cfs_rq);
2358 check_enqueue_throttle(cfs_rq);
2362 static void __clear_buddies_last(struct sched_entity *se)
2364 for_each_sched_entity(se) {
2365 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2366 if (cfs_rq->last == se)
2367 cfs_rq->last = NULL;
2373 static void __clear_buddies_next(struct sched_entity *se)
2375 for_each_sched_entity(se) {
2376 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2377 if (cfs_rq->next == se)
2378 cfs_rq->next = NULL;
2384 static void __clear_buddies_skip(struct sched_entity *se)
2386 for_each_sched_entity(se) {
2387 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2388 if (cfs_rq->skip == se)
2389 cfs_rq->skip = NULL;
2395 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2397 if (cfs_rq->last == se)
2398 __clear_buddies_last(se);
2400 if (cfs_rq->next == se)
2401 __clear_buddies_next(se);
2403 if (cfs_rq->skip == se)
2404 __clear_buddies_skip(se);
2407 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2410 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2413 * Update run-time statistics of the 'current'.
2415 update_curr(cfs_rq);
2416 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2418 update_stats_dequeue(cfs_rq, se);
2419 if (flags & DEQUEUE_SLEEP) {
2420 #ifdef CONFIG_SCHEDSTATS
2421 if (entity_is_task(se)) {
2422 struct task_struct *tsk = task_of(se);
2424 if (tsk->state & TASK_INTERRUPTIBLE)
2425 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2426 if (tsk->state & TASK_UNINTERRUPTIBLE)
2427 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2432 clear_buddies(cfs_rq, se);
2434 if (se != cfs_rq->curr)
2435 __dequeue_entity(cfs_rq, se);
2437 account_entity_dequeue(cfs_rq, se);
2440 * Normalize the entity after updating the min_vruntime because the
2441 * update can refer to the ->curr item and we need to reflect this
2442 * movement in our normalized position.
2444 if (!(flags & DEQUEUE_SLEEP))
2445 se->vruntime -= cfs_rq->min_vruntime;
2447 /* return excess runtime on last dequeue */
2448 return_cfs_rq_runtime(cfs_rq);
2450 update_min_vruntime(cfs_rq);
2451 update_cfs_shares(cfs_rq);
2455 * Preempt the current task with a newly woken task if needed:
2458 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2460 unsigned long ideal_runtime, delta_exec;
2461 struct sched_entity *se;
2464 ideal_runtime = sched_slice(cfs_rq, curr);
2465 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2466 if (delta_exec > ideal_runtime) {
2467 resched_task(rq_of(cfs_rq)->curr);
2469 * The current task ran long enough, ensure it doesn't get
2470 * re-elected due to buddy favours.
2472 clear_buddies(cfs_rq, curr);
2477 * Ensure that a task that missed wakeup preemption by a
2478 * narrow margin doesn't have to wait for a full slice.
2479 * This also mitigates buddy induced latencies under load.
2481 if (delta_exec < sysctl_sched_min_granularity)
2484 se = __pick_first_entity(cfs_rq);
2485 delta = curr->vruntime - se->vruntime;
2490 if (delta > ideal_runtime)
2491 resched_task(rq_of(cfs_rq)->curr);
2495 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2497 /* 'current' is not kept within the tree. */
2500 * Any task has to be enqueued before it get to execute on
2501 * a CPU. So account for the time it spent waiting on the
2504 update_stats_wait_end(cfs_rq, se);
2505 __dequeue_entity(cfs_rq, se);
2508 update_stats_curr_start(cfs_rq, se);
2510 #ifdef CONFIG_SCHEDSTATS
2512 * Track our maximum slice length, if the CPU's load is at
2513 * least twice that of our own weight (i.e. dont track it
2514 * when there are only lesser-weight tasks around):
2516 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2517 se->statistics.slice_max = max(se->statistics.slice_max,
2518 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2521 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2525 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2528 * Pick the next process, keeping these things in mind, in this order:
2529 * 1) keep things fair between processes/task groups
2530 * 2) pick the "next" process, since someone really wants that to run
2531 * 3) pick the "last" process, for cache locality
2532 * 4) do not run the "skip" process, if something else is available
2534 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2536 struct sched_entity *se = __pick_first_entity(cfs_rq);
2537 struct sched_entity *left = se;
2540 * Avoid running the skip buddy, if running something else can
2541 * be done without getting too unfair.
2543 if (cfs_rq->skip == se) {
2544 struct sched_entity *second = __pick_next_entity(se);
2545 if (second && wakeup_preempt_entity(second, left) < 1)
2550 * Prefer last buddy, try to return the CPU to a preempted task.
2552 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2556 * Someone really wants this to run. If it's not unfair, run it.
2558 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2561 clear_buddies(cfs_rq, se);
2566 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2568 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2571 * If still on the runqueue then deactivate_task()
2572 * was not called and update_curr() has to be done:
2575 update_curr(cfs_rq);
2577 /* throttle cfs_rqs exceeding runtime */
2578 check_cfs_rq_runtime(cfs_rq);
2580 check_spread(cfs_rq, prev);
2582 update_stats_wait_start(cfs_rq, prev);
2583 /* Put 'current' back into the tree. */
2584 __enqueue_entity(cfs_rq, prev);
2585 /* in !on_rq case, update occurred at dequeue */
2586 update_entity_load_avg(prev, 1);
2588 cfs_rq->curr = NULL;
2592 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2595 * Update run-time statistics of the 'current'.
2597 update_curr(cfs_rq);
2600 * Ensure that runnable average is periodically updated.
2602 update_entity_load_avg(curr, 1);
2603 update_cfs_rq_blocked_load(cfs_rq, 1);
2604 update_cfs_shares(cfs_rq);
2606 #ifdef CONFIG_SCHED_HRTICK
2608 * queued ticks are scheduled to match the slice, so don't bother
2609 * validating it and just reschedule.
2612 resched_task(rq_of(cfs_rq)->curr);
2616 * don't let the period tick interfere with the hrtick preemption
2618 if (!sched_feat(DOUBLE_TICK) &&
2619 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2623 if (cfs_rq->nr_running > 1)
2624 check_preempt_tick(cfs_rq, curr);
2628 /**************************************************
2629 * CFS bandwidth control machinery
2632 #ifdef CONFIG_CFS_BANDWIDTH
2634 #ifdef HAVE_JUMP_LABEL
2635 static struct static_key __cfs_bandwidth_used;
2637 static inline bool cfs_bandwidth_used(void)
2639 return static_key_false(&__cfs_bandwidth_used);
2642 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2644 /* only need to count groups transitioning between enabled/!enabled */
2645 if (enabled && !was_enabled)
2646 static_key_slow_inc(&__cfs_bandwidth_used);
2647 else if (!enabled && was_enabled)
2648 static_key_slow_dec(&__cfs_bandwidth_used);
2650 #else /* HAVE_JUMP_LABEL */
2651 static bool cfs_bandwidth_used(void)
2656 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2657 #endif /* HAVE_JUMP_LABEL */
2660 * default period for cfs group bandwidth.
2661 * default: 0.1s, units: nanoseconds
2663 static inline u64 default_cfs_period(void)
2665 return 100000000ULL;
2668 static inline u64 sched_cfs_bandwidth_slice(void)
2670 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2674 * Replenish runtime according to assigned quota and update expiration time.
2675 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2676 * additional synchronization around rq->lock.
2678 * requires cfs_b->lock
2680 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2684 if (cfs_b->quota == RUNTIME_INF)
2687 now = sched_clock_cpu(smp_processor_id());
2688 cfs_b->runtime = cfs_b->quota;
2689 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2692 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2694 return &tg->cfs_bandwidth;
2697 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2698 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2700 if (unlikely(cfs_rq->throttle_count))
2701 return cfs_rq->throttled_clock_task;
2703 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2706 /* returns 0 on failure to allocate runtime */
2707 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2709 struct task_group *tg = cfs_rq->tg;
2710 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2711 u64 amount = 0, min_amount, expires;
2713 /* note: this is a positive sum as runtime_remaining <= 0 */
2714 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2716 raw_spin_lock(&cfs_b->lock);
2717 if (cfs_b->quota == RUNTIME_INF)
2718 amount = min_amount;
2721 * If the bandwidth pool has become inactive, then at least one
2722 * period must have elapsed since the last consumption.
2723 * Refresh the global state and ensure bandwidth timer becomes
2726 if (!cfs_b->timer_active) {
2727 __refill_cfs_bandwidth_runtime(cfs_b);
2728 __start_cfs_bandwidth(cfs_b);
2731 if (cfs_b->runtime > 0) {
2732 amount = min(cfs_b->runtime, min_amount);
2733 cfs_b->runtime -= amount;
2737 expires = cfs_b->runtime_expires;
2738 raw_spin_unlock(&cfs_b->lock);
2740 cfs_rq->runtime_remaining += amount;
2742 * we may have advanced our local expiration to account for allowed
2743 * spread between our sched_clock and the one on which runtime was
2746 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2747 cfs_rq->runtime_expires = expires;
2749 return cfs_rq->runtime_remaining > 0;
2753 * Note: This depends on the synchronization provided by sched_clock and the
2754 * fact that rq->clock snapshots this value.
2756 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2758 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2760 /* if the deadline is ahead of our clock, nothing to do */
2761 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2764 if (cfs_rq->runtime_remaining < 0)
2768 * If the local deadline has passed we have to consider the
2769 * possibility that our sched_clock is 'fast' and the global deadline
2770 * has not truly expired.
2772 * Fortunately we can check determine whether this the case by checking
2773 * whether the global deadline has advanced.
2776 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2777 /* extend local deadline, drift is bounded above by 2 ticks */
2778 cfs_rq->runtime_expires += TICK_NSEC;
2780 /* global deadline is ahead, expiration has passed */
2781 cfs_rq->runtime_remaining = 0;
2785 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2786 unsigned long delta_exec)
2788 /* dock delta_exec before expiring quota (as it could span periods) */
2789 cfs_rq->runtime_remaining -= delta_exec;
2790 expire_cfs_rq_runtime(cfs_rq);
2792 if (likely(cfs_rq->runtime_remaining > 0))
2796 * if we're unable to extend our runtime we resched so that the active
2797 * hierarchy can be throttled
2799 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2800 resched_task(rq_of(cfs_rq)->curr);
2803 static __always_inline
2804 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2806 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2809 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2812 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2814 return cfs_bandwidth_used() && cfs_rq->throttled;
2817 /* check whether cfs_rq, or any parent, is throttled */
2818 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2820 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2824 * Ensure that neither of the group entities corresponding to src_cpu or
2825 * dest_cpu are members of a throttled hierarchy when performing group
2826 * load-balance operations.
2828 static inline int throttled_lb_pair(struct task_group *tg,
2829 int src_cpu, int dest_cpu)
2831 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2833 src_cfs_rq = tg->cfs_rq[src_cpu];
2834 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2836 return throttled_hierarchy(src_cfs_rq) ||
2837 throttled_hierarchy(dest_cfs_rq);
2840 /* updated child weight may affect parent so we have to do this bottom up */
2841 static int tg_unthrottle_up(struct task_group *tg, void *data)
2843 struct rq *rq = data;
2844 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2846 cfs_rq->throttle_count--;
2848 if (!cfs_rq->throttle_count) {
2849 /* adjust cfs_rq_clock_task() */
2850 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2851 cfs_rq->throttled_clock_task;
2858 static int tg_throttle_down(struct task_group *tg, void *data)
2860 struct rq *rq = data;
2861 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2863 /* group is entering throttled state, stop time */
2864 if (!cfs_rq->throttle_count)
2865 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2866 cfs_rq->throttle_count++;
2871 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2873 struct rq *rq = rq_of(cfs_rq);
2874 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2875 struct sched_entity *se;
2876 long task_delta, dequeue = 1;
2878 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2880 /* freeze hierarchy runnable averages while throttled */
2882 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2885 task_delta = cfs_rq->h_nr_running;
2886 for_each_sched_entity(se) {
2887 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2888 /* throttled entity or throttle-on-deactivate */
2893 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2894 qcfs_rq->h_nr_running -= task_delta;
2896 if (qcfs_rq->load.weight)
2901 rq->nr_running -= task_delta;
2903 cfs_rq->throttled = 1;
2904 cfs_rq->throttled_clock = rq_clock(rq);
2905 raw_spin_lock(&cfs_b->lock);
2906 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2907 raw_spin_unlock(&cfs_b->lock);
2910 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2912 struct rq *rq = rq_of(cfs_rq);
2913 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2914 struct sched_entity *se;
2918 se = cfs_rq->tg->se[cpu_of(rq)];
2920 cfs_rq->throttled = 0;
2922 update_rq_clock(rq);
2924 raw_spin_lock(&cfs_b->lock);
2925 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2926 list_del_rcu(&cfs_rq->throttled_list);
2927 raw_spin_unlock(&cfs_b->lock);
2929 /* update hierarchical throttle state */
2930 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2932 if (!cfs_rq->load.weight)
2935 task_delta = cfs_rq->h_nr_running;
2936 for_each_sched_entity(se) {
2940 cfs_rq = cfs_rq_of(se);
2942 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2943 cfs_rq->h_nr_running += task_delta;
2945 if (cfs_rq_throttled(cfs_rq))
2950 rq->nr_running += task_delta;
2952 /* determine whether we need to wake up potentially idle cpu */
2953 if (rq->curr == rq->idle && rq->cfs.nr_running)
2954 resched_task(rq->curr);
2957 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2958 u64 remaining, u64 expires)
2960 struct cfs_rq *cfs_rq;
2961 u64 runtime = remaining;
2964 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2966 struct rq *rq = rq_of(cfs_rq);
2968 raw_spin_lock(&rq->lock);
2969 if (!cfs_rq_throttled(cfs_rq))
2972 runtime = -cfs_rq->runtime_remaining + 1;
2973 if (runtime > remaining)
2974 runtime = remaining;
2975 remaining -= runtime;
2977 cfs_rq->runtime_remaining += runtime;
2978 cfs_rq->runtime_expires = expires;
2980 /* we check whether we're throttled above */
2981 if (cfs_rq->runtime_remaining > 0)
2982 unthrottle_cfs_rq(cfs_rq);
2985 raw_spin_unlock(&rq->lock);
2996 * Responsible for refilling a task_group's bandwidth and unthrottling its
2997 * cfs_rqs as appropriate. If there has been no activity within the last
2998 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2999 * used to track this state.
3001 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3003 u64 runtime, runtime_expires;
3004 int idle = 1, throttled;
3006 raw_spin_lock(&cfs_b->lock);
3007 /* no need to continue the timer with no bandwidth constraint */
3008 if (cfs_b->quota == RUNTIME_INF)
3011 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3012 /* idle depends on !throttled (for the case of a large deficit) */
3013 idle = cfs_b->idle && !throttled;
3014 cfs_b->nr_periods += overrun;
3016 /* if we're going inactive then everything else can be deferred */
3020 __refill_cfs_bandwidth_runtime(cfs_b);
3023 /* mark as potentially idle for the upcoming period */
3028 /* account preceding periods in which throttling occurred */
3029 cfs_b->nr_throttled += overrun;
3032 * There are throttled entities so we must first use the new bandwidth
3033 * to unthrottle them before making it generally available. This
3034 * ensures that all existing debts will be paid before a new cfs_rq is
3037 runtime = cfs_b->runtime;
3038 runtime_expires = cfs_b->runtime_expires;
3042 * This check is repeated as we are holding onto the new bandwidth
3043 * while we unthrottle. This can potentially race with an unthrottled
3044 * group trying to acquire new bandwidth from the global pool.
3046 while (throttled && runtime > 0) {
3047 raw_spin_unlock(&cfs_b->lock);
3048 /* we can't nest cfs_b->lock while distributing bandwidth */
3049 runtime = distribute_cfs_runtime(cfs_b, runtime,
3051 raw_spin_lock(&cfs_b->lock);
3053 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3056 /* return (any) remaining runtime */
3057 cfs_b->runtime = runtime;
3059 * While we are ensured activity in the period following an
3060 * unthrottle, this also covers the case in which the new bandwidth is
3061 * insufficient to cover the existing bandwidth deficit. (Forcing the
3062 * timer to remain active while there are any throttled entities.)
3067 cfs_b->timer_active = 0;
3068 raw_spin_unlock(&cfs_b->lock);
3073 /* a cfs_rq won't donate quota below this amount */
3074 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3075 /* minimum remaining period time to redistribute slack quota */
3076 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3077 /* how long we wait to gather additional slack before distributing */
3078 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3080 /* are we near the end of the current quota period? */
3081 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3083 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3086 /* if the call-back is running a quota refresh is already occurring */
3087 if (hrtimer_callback_running(refresh_timer))
3090 /* is a quota refresh about to occur? */
3091 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3092 if (remaining < min_expire)
3098 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3100 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3102 /* if there's a quota refresh soon don't bother with slack */
3103 if (runtime_refresh_within(cfs_b, min_left))
3106 start_bandwidth_timer(&cfs_b->slack_timer,
3107 ns_to_ktime(cfs_bandwidth_slack_period));
3110 /* we know any runtime found here is valid as update_curr() precedes return */
3111 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3113 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3114 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3116 if (slack_runtime <= 0)
3119 raw_spin_lock(&cfs_b->lock);
3120 if (cfs_b->quota != RUNTIME_INF &&
3121 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3122 cfs_b->runtime += slack_runtime;
3124 /* we are under rq->lock, defer unthrottling using a timer */
3125 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3126 !list_empty(&cfs_b->throttled_cfs_rq))
3127 start_cfs_slack_bandwidth(cfs_b);
3129 raw_spin_unlock(&cfs_b->lock);
3131 /* even if it's not valid for return we don't want to try again */
3132 cfs_rq->runtime_remaining -= slack_runtime;
3135 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3137 if (!cfs_bandwidth_used())
3140 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3143 __return_cfs_rq_runtime(cfs_rq);
3147 * This is done with a timer (instead of inline with bandwidth return) since
3148 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3150 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3152 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3155 /* confirm we're still not at a refresh boundary */
3156 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3159 raw_spin_lock(&cfs_b->lock);
3160 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3161 runtime = cfs_b->runtime;
3164 expires = cfs_b->runtime_expires;
3165 raw_spin_unlock(&cfs_b->lock);
3170 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3172 raw_spin_lock(&cfs_b->lock);
3173 if (expires == cfs_b->runtime_expires)
3174 cfs_b->runtime = runtime;
3175 raw_spin_unlock(&cfs_b->lock);
3179 * When a group wakes up we want to make sure that its quota is not already
3180 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3181 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3183 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3185 if (!cfs_bandwidth_used())
3188 /* an active group must be handled by the update_curr()->put() path */
3189 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3192 /* ensure the group is not already throttled */
3193 if (cfs_rq_throttled(cfs_rq))
3196 /* update runtime allocation */
3197 account_cfs_rq_runtime(cfs_rq, 0);
3198 if (cfs_rq->runtime_remaining <= 0)
3199 throttle_cfs_rq(cfs_rq);
3202 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3203 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3205 if (!cfs_bandwidth_used())
3208 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3212 * it's possible for a throttled entity to be forced into a running
3213 * state (e.g. set_curr_task), in this case we're finished.
3215 if (cfs_rq_throttled(cfs_rq))
3218 throttle_cfs_rq(cfs_rq);
3221 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3223 struct cfs_bandwidth *cfs_b =
3224 container_of(timer, struct cfs_bandwidth, slack_timer);
3225 do_sched_cfs_slack_timer(cfs_b);
3227 return HRTIMER_NORESTART;
3230 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3232 struct cfs_bandwidth *cfs_b =
3233 container_of(timer, struct cfs_bandwidth, period_timer);
3239 now = hrtimer_cb_get_time(timer);
3240 overrun = hrtimer_forward(timer, now, cfs_b->period);
3245 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3248 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3251 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3253 raw_spin_lock_init(&cfs_b->lock);
3255 cfs_b->quota = RUNTIME_INF;
3256 cfs_b->period = ns_to_ktime(default_cfs_period());
3258 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3259 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3260 cfs_b->period_timer.function = sched_cfs_period_timer;
3261 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3262 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3265 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3267 cfs_rq->runtime_enabled = 0;
3268 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3271 /* requires cfs_b->lock, may release to reprogram timer */
3272 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3275 * The timer may be active because we're trying to set a new bandwidth
3276 * period or because we're racing with the tear-down path
3277 * (timer_active==0 becomes visible before the hrtimer call-back
3278 * terminates). In either case we ensure that it's re-programmed
3280 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3281 raw_spin_unlock(&cfs_b->lock);
3282 /* ensure cfs_b->lock is available while we wait */
3283 hrtimer_cancel(&cfs_b->period_timer);
3285 raw_spin_lock(&cfs_b->lock);
3286 /* if someone else restarted the timer then we're done */
3287 if (cfs_b->timer_active)
3291 cfs_b->timer_active = 1;
3292 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3295 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3297 hrtimer_cancel(&cfs_b->period_timer);
3298 hrtimer_cancel(&cfs_b->slack_timer);
3301 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3303 struct cfs_rq *cfs_rq;
3305 for_each_leaf_cfs_rq(rq, cfs_rq) {
3306 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3308 if (!cfs_rq->runtime_enabled)
3312 * clock_task is not advancing so we just need to make sure
3313 * there's some valid quota amount
3315 cfs_rq->runtime_remaining = cfs_b->quota;
3316 if (cfs_rq_throttled(cfs_rq))
3317 unthrottle_cfs_rq(cfs_rq);
3321 #else /* CONFIG_CFS_BANDWIDTH */
3322 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3324 return rq_clock_task(rq_of(cfs_rq));
3327 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3328 unsigned long delta_exec) {}
3329 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3330 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3331 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3333 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3338 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3343 static inline int throttled_lb_pair(struct task_group *tg,
3344 int src_cpu, int dest_cpu)
3349 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3351 #ifdef CONFIG_FAIR_GROUP_SCHED
3352 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3355 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3359 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3360 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3362 #endif /* CONFIG_CFS_BANDWIDTH */
3364 /**************************************************
3365 * CFS operations on tasks:
3368 #ifdef CONFIG_SCHED_HRTICK
3369 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3371 struct sched_entity *se = &p->se;
3372 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3374 WARN_ON(task_rq(p) != rq);
3376 if (cfs_rq->nr_running > 1) {
3377 u64 slice = sched_slice(cfs_rq, se);
3378 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3379 s64 delta = slice - ran;
3388 * Don't schedule slices shorter than 10000ns, that just
3389 * doesn't make sense. Rely on vruntime for fairness.
3392 delta = max_t(s64, 10000LL, delta);
3394 hrtick_start(rq, delta);
3399 * called from enqueue/dequeue and updates the hrtick when the
3400 * current task is from our class and nr_running is low enough
3403 static void hrtick_update(struct rq *rq)
3405 struct task_struct *curr = rq->curr;
3407 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3410 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3411 hrtick_start_fair(rq, curr);
3413 #else /* !CONFIG_SCHED_HRTICK */
3415 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3419 static inline void hrtick_update(struct rq *rq)
3425 * The enqueue_task method is called before nr_running is
3426 * increased. Here we update the fair scheduling stats and
3427 * then put the task into the rbtree:
3430 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3432 struct cfs_rq *cfs_rq;
3433 struct sched_entity *se = &p->se;
3435 for_each_sched_entity(se) {
3438 cfs_rq = cfs_rq_of(se);
3439 enqueue_entity(cfs_rq, se, flags);
3442 * end evaluation on encountering a throttled cfs_rq
3444 * note: in the case of encountering a throttled cfs_rq we will
3445 * post the final h_nr_running increment below.
3447 if (cfs_rq_throttled(cfs_rq))
3449 cfs_rq->h_nr_running++;
3451 flags = ENQUEUE_WAKEUP;
3454 for_each_sched_entity(se) {
3455 cfs_rq = cfs_rq_of(se);
3456 cfs_rq->h_nr_running++;
3458 if (cfs_rq_throttled(cfs_rq))
3461 update_cfs_shares(cfs_rq);
3462 update_entity_load_avg(se, 1);
3466 update_rq_runnable_avg(rq, rq->nr_running);
3472 static void set_next_buddy(struct sched_entity *se);
3475 * The dequeue_task method is called before nr_running is
3476 * decreased. We remove the task from the rbtree and
3477 * update the fair scheduling stats:
3479 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3481 struct cfs_rq *cfs_rq;
3482 struct sched_entity *se = &p->se;
3483 int task_sleep = flags & DEQUEUE_SLEEP;
3485 for_each_sched_entity(se) {
3486 cfs_rq = cfs_rq_of(se);
3487 dequeue_entity(cfs_rq, se, flags);
3490 * end evaluation on encountering a throttled cfs_rq
3492 * note: in the case of encountering a throttled cfs_rq we will
3493 * post the final h_nr_running decrement below.
3495 if (cfs_rq_throttled(cfs_rq))
3497 cfs_rq->h_nr_running--;
3499 /* Don't dequeue parent if it has other entities besides us */
3500 if (cfs_rq->load.weight) {
3502 * Bias pick_next to pick a task from this cfs_rq, as
3503 * p is sleeping when it is within its sched_slice.
3505 if (task_sleep && parent_entity(se))
3506 set_next_buddy(parent_entity(se));
3508 /* avoid re-evaluating load for this entity */
3509 se = parent_entity(se);
3512 flags |= DEQUEUE_SLEEP;
3515 for_each_sched_entity(se) {
3516 cfs_rq = cfs_rq_of(se);
3517 cfs_rq->h_nr_running--;
3519 if (cfs_rq_throttled(cfs_rq))
3522 update_cfs_shares(cfs_rq);
3523 update_entity_load_avg(se, 1);
3528 update_rq_runnable_avg(rq, 1);
3534 /* Used instead of source_load when we know the type == 0 */
3535 static unsigned long weighted_cpuload(const int cpu)
3537 return cpu_rq(cpu)->cfs.runnable_load_avg;
3541 * Return a low guess at the load of a migration-source cpu weighted
3542 * according to the scheduling class and "nice" value.
3544 * We want to under-estimate the load of migration sources, to
3545 * balance conservatively.
3547 static unsigned long source_load(int cpu, int type)
3549 struct rq *rq = cpu_rq(cpu);
3550 unsigned long total = weighted_cpuload(cpu);
3552 if (type == 0 || !sched_feat(LB_BIAS))
3555 return min(rq->cpu_load[type-1], total);
3559 * Return a high guess at the load of a migration-target cpu weighted
3560 * according to the scheduling class and "nice" value.
3562 static unsigned long target_load(int cpu, int type)
3564 struct rq *rq = cpu_rq(cpu);
3565 unsigned long total = weighted_cpuload(cpu);
3567 if (type == 0 || !sched_feat(LB_BIAS))
3570 return max(rq->cpu_load[type-1], total);
3573 static unsigned long power_of(int cpu)
3575 return cpu_rq(cpu)->cpu_power;
3578 static unsigned long cpu_avg_load_per_task(int cpu)
3580 struct rq *rq = cpu_rq(cpu);
3581 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3582 unsigned long load_avg = rq->cfs.runnable_load_avg;
3585 return load_avg / nr_running;
3590 static void record_wakee(struct task_struct *p)
3593 * Rough decay (wiping) for cost saving, don't worry
3594 * about the boundary, really active task won't care
3597 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3598 current->wakee_flips = 0;
3599 current->wakee_flip_decay_ts = jiffies;
3602 if (current->last_wakee != p) {
3603 current->last_wakee = p;
3604 current->wakee_flips++;
3608 static void task_waking_fair(struct task_struct *p)
3610 struct sched_entity *se = &p->se;
3611 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3614 #ifndef CONFIG_64BIT
3615 u64 min_vruntime_copy;
3618 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3620 min_vruntime = cfs_rq->min_vruntime;
3621 } while (min_vruntime != min_vruntime_copy);
3623 min_vruntime = cfs_rq->min_vruntime;
3626 se->vruntime -= min_vruntime;
3630 #ifdef CONFIG_FAIR_GROUP_SCHED
3632 * effective_load() calculates the load change as seen from the root_task_group
3634 * Adding load to a group doesn't make a group heavier, but can cause movement
3635 * of group shares between cpus. Assuming the shares were perfectly aligned one
3636 * can calculate the shift in shares.
3638 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3639 * on this @cpu and results in a total addition (subtraction) of @wg to the
3640 * total group weight.
3642 * Given a runqueue weight distribution (rw_i) we can compute a shares
3643 * distribution (s_i) using:
3645 * s_i = rw_i / \Sum rw_j (1)
3647 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3648 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3649 * shares distribution (s_i):
3651 * rw_i = { 2, 4, 1, 0 }
3652 * s_i = { 2/7, 4/7, 1/7, 0 }
3654 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3655 * task used to run on and the CPU the waker is running on), we need to
3656 * compute the effect of waking a task on either CPU and, in case of a sync
3657 * wakeup, compute the effect of the current task going to sleep.
3659 * So for a change of @wl to the local @cpu with an overall group weight change
3660 * of @wl we can compute the new shares distribution (s'_i) using:
3662 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3664 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3665 * differences in waking a task to CPU 0. The additional task changes the
3666 * weight and shares distributions like:
3668 * rw'_i = { 3, 4, 1, 0 }
3669 * s'_i = { 3/8, 4/8, 1/8, 0 }
3671 * We can then compute the difference in effective weight by using:
3673 * dw_i = S * (s'_i - s_i) (3)
3675 * Where 'S' is the group weight as seen by its parent.
3677 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3678 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3679 * 4/7) times the weight of the group.
3681 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3683 struct sched_entity *se = tg->se[cpu];
3685 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3688 for_each_sched_entity(se) {
3694 * W = @wg + \Sum rw_j
3696 W = wg + calc_tg_weight(tg, se->my_q);
3701 w = se->my_q->load.weight + wl;
3704 * wl = S * s'_i; see (2)
3707 wl = (w * tg->shares) / W;
3712 * Per the above, wl is the new se->load.weight value; since
3713 * those are clipped to [MIN_SHARES, ...) do so now. See
3714 * calc_cfs_shares().
3716 if (wl < MIN_SHARES)
3720 * wl = dw_i = S * (s'_i - s_i); see (3)
3722 wl -= se->load.weight;
3725 * Recursively apply this logic to all parent groups to compute
3726 * the final effective load change on the root group. Since
3727 * only the @tg group gets extra weight, all parent groups can
3728 * only redistribute existing shares. @wl is the shift in shares
3729 * resulting from this level per the above.
3738 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3745 static int wake_wide(struct task_struct *p)
3747 int factor = this_cpu_read(sd_llc_size);
3750 * Yeah, it's the switching-frequency, could means many wakee or
3751 * rapidly switch, use factor here will just help to automatically
3752 * adjust the loose-degree, so bigger node will lead to more pull.
3754 if (p->wakee_flips > factor) {
3756 * wakee is somewhat hot, it needs certain amount of cpu
3757 * resource, so if waker is far more hot, prefer to leave
3760 if (current->wakee_flips > (factor * p->wakee_flips))
3767 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3769 s64 this_load, load;
3770 int idx, this_cpu, prev_cpu;
3771 unsigned long tl_per_task;
3772 struct task_group *tg;
3773 unsigned long weight;
3777 * If we wake multiple tasks be careful to not bounce
3778 * ourselves around too much.
3784 this_cpu = smp_processor_id();
3785 prev_cpu = task_cpu(p);
3786 load = source_load(prev_cpu, idx);
3787 this_load = target_load(this_cpu, idx);
3790 * If sync wakeup then subtract the (maximum possible)
3791 * effect of the currently running task from the load
3792 * of the current CPU:
3795 tg = task_group(current);
3796 weight = current->se.load.weight;
3798 this_load += effective_load(tg, this_cpu, -weight, -weight);
3799 load += effective_load(tg, prev_cpu, 0, -weight);
3803 weight = p->se.load.weight;
3806 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3807 * due to the sync cause above having dropped this_load to 0, we'll
3808 * always have an imbalance, but there's really nothing you can do
3809 * about that, so that's good too.
3811 * Otherwise check if either cpus are near enough in load to allow this
3812 * task to be woken on this_cpu.
3814 if (this_load > 0) {
3815 s64 this_eff_load, prev_eff_load;
3817 this_eff_load = 100;
3818 this_eff_load *= power_of(prev_cpu);
3819 this_eff_load *= this_load +
3820 effective_load(tg, this_cpu, weight, weight);
3822 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3823 prev_eff_load *= power_of(this_cpu);
3824 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3826 balanced = this_eff_load <= prev_eff_load;
3831 * If the currently running task will sleep within
3832 * a reasonable amount of time then attract this newly
3835 if (sync && balanced)
3838 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3839 tl_per_task = cpu_avg_load_per_task(this_cpu);
3842 (this_load <= load &&
3843 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3845 * This domain has SD_WAKE_AFFINE and
3846 * p is cache cold in this domain, and
3847 * there is no bad imbalance.
3849 schedstat_inc(sd, ttwu_move_affine);
3850 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3858 * find_idlest_group finds and returns the least busy CPU group within the
3861 static struct sched_group *
3862 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3863 int this_cpu, int load_idx)
3865 struct sched_group *idlest = NULL, *group = sd->groups;
3866 unsigned long min_load = ULONG_MAX, this_load = 0;
3867 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3870 unsigned long load, avg_load;
3874 /* Skip over this group if it has no CPUs allowed */
3875 if (!cpumask_intersects(sched_group_cpus(group),
3876 tsk_cpus_allowed(p)))
3879 local_group = cpumask_test_cpu(this_cpu,
3880 sched_group_cpus(group));
3882 /* Tally up the load of all CPUs in the group */
3885 for_each_cpu(i, sched_group_cpus(group)) {
3886 /* Bias balancing toward cpus of our domain */
3888 load = source_load(i, load_idx);
3890 load = target_load(i, load_idx);
3895 /* Adjust by relative CPU power of the group */
3896 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3899 this_load = avg_load;
3900 } else if (avg_load < min_load) {
3901 min_load = avg_load;
3904 } while (group = group->next, group != sd->groups);
3906 if (!idlest || 100*this_load < imbalance*min_load)
3912 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3915 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3917 unsigned long load, min_load = ULONG_MAX;
3921 /* Traverse only the allowed CPUs */
3922 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3923 load = weighted_cpuload(i);
3925 if (load < min_load || (load == min_load && i == this_cpu)) {
3935 * Try and locate an idle CPU in the sched_domain.
3937 static int select_idle_sibling(struct task_struct *p, int target)
3939 struct sched_domain *sd;
3940 struct sched_group *sg;
3941 int i = task_cpu(p);
3943 if (idle_cpu(target))
3947 * If the prevous cpu is cache affine and idle, don't be stupid.
3949 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3953 * Otherwise, iterate the domains and find an elegible idle cpu.
3955 sd = rcu_dereference(per_cpu(sd_llc, target));
3956 for_each_lower_domain(sd) {
3959 if (!cpumask_intersects(sched_group_cpus(sg),
3960 tsk_cpus_allowed(p)))
3963 for_each_cpu(i, sched_group_cpus(sg)) {
3964 if (i == target || !idle_cpu(i))
3968 target = cpumask_first_and(sched_group_cpus(sg),
3969 tsk_cpus_allowed(p));
3973 } while (sg != sd->groups);
3980 * sched_balance_self: balance the current task (running on cpu) in domains
3981 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3984 * Balance, ie. select the least loaded group.
3986 * Returns the target CPU number, or the same CPU if no balancing is needed.
3988 * preempt must be disabled.
3991 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
3993 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3994 int cpu = smp_processor_id();
3996 int want_affine = 0;
3997 int sync = wake_flags & WF_SYNC;
3999 if (p->nr_cpus_allowed == 1)
4002 if (sd_flag & SD_BALANCE_WAKE) {
4003 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4009 for_each_domain(cpu, tmp) {
4010 if (!(tmp->flags & SD_LOAD_BALANCE))
4014 * If both cpu and prev_cpu are part of this domain,
4015 * cpu is a valid SD_WAKE_AFFINE target.
4017 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4018 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4023 if (tmp->flags & sd_flag)
4028 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4031 new_cpu = select_idle_sibling(p, prev_cpu);
4036 int load_idx = sd->forkexec_idx;
4037 struct sched_group *group;
4040 if (!(sd->flags & sd_flag)) {
4045 if (sd_flag & SD_BALANCE_WAKE)
4046 load_idx = sd->wake_idx;
4048 group = find_idlest_group(sd, p, cpu, load_idx);
4054 new_cpu = find_idlest_cpu(group, p, cpu);
4055 if (new_cpu == -1 || new_cpu == cpu) {
4056 /* Now try balancing at a lower domain level of cpu */
4061 /* Now try balancing at a lower domain level of new_cpu */
4063 weight = sd->span_weight;
4065 for_each_domain(cpu, tmp) {
4066 if (weight <= tmp->span_weight)
4068 if (tmp->flags & sd_flag)
4071 /* while loop will break here if sd == NULL */
4080 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4081 * cfs_rq_of(p) references at time of call are still valid and identify the
4082 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4083 * other assumptions, including the state of rq->lock, should be made.
4086 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4088 struct sched_entity *se = &p->se;
4089 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4092 * Load tracking: accumulate removed load so that it can be processed
4093 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4094 * to blocked load iff they have a positive decay-count. It can never
4095 * be negative here since on-rq tasks have decay-count == 0.
4097 if (se->avg.decay_count) {
4098 se->avg.decay_count = -__synchronize_entity_decay(se);
4099 atomic_long_add(se->avg.load_avg_contrib,
4100 &cfs_rq->removed_load);
4103 #endif /* CONFIG_SMP */
4105 static unsigned long
4106 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4108 unsigned long gran = sysctl_sched_wakeup_granularity;
4111 * Since its curr running now, convert the gran from real-time
4112 * to virtual-time in his units.
4114 * By using 'se' instead of 'curr' we penalize light tasks, so
4115 * they get preempted easier. That is, if 'se' < 'curr' then
4116 * the resulting gran will be larger, therefore penalizing the
4117 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4118 * be smaller, again penalizing the lighter task.
4120 * This is especially important for buddies when the leftmost
4121 * task is higher priority than the buddy.
4123 return calc_delta_fair(gran, se);
4127 * Should 'se' preempt 'curr'.
4141 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4143 s64 gran, vdiff = curr->vruntime - se->vruntime;
4148 gran = wakeup_gran(curr, se);
4155 static void set_last_buddy(struct sched_entity *se)
4157 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4160 for_each_sched_entity(se)
4161 cfs_rq_of(se)->last = se;
4164 static void set_next_buddy(struct sched_entity *se)
4166 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4169 for_each_sched_entity(se)
4170 cfs_rq_of(se)->next = se;
4173 static void set_skip_buddy(struct sched_entity *se)
4175 for_each_sched_entity(se)
4176 cfs_rq_of(se)->skip = se;
4180 * Preempt the current task with a newly woken task if needed:
4182 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4184 struct task_struct *curr = rq->curr;
4185 struct sched_entity *se = &curr->se, *pse = &p->se;
4186 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4187 int scale = cfs_rq->nr_running >= sched_nr_latency;
4188 int next_buddy_marked = 0;
4190 if (unlikely(se == pse))
4194 * This is possible from callers such as move_task(), in which we
4195 * unconditionally check_prempt_curr() after an enqueue (which may have
4196 * lead to a throttle). This both saves work and prevents false
4197 * next-buddy nomination below.
4199 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4202 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4203 set_next_buddy(pse);
4204 next_buddy_marked = 1;
4208 * We can come here with TIF_NEED_RESCHED already set from new task
4211 * Note: this also catches the edge-case of curr being in a throttled
4212 * group (e.g. via set_curr_task), since update_curr() (in the
4213 * enqueue of curr) will have resulted in resched being set. This
4214 * prevents us from potentially nominating it as a false LAST_BUDDY
4217 if (test_tsk_need_resched(curr))
4220 /* Idle tasks are by definition preempted by non-idle tasks. */
4221 if (unlikely(curr->policy == SCHED_IDLE) &&
4222 likely(p->policy != SCHED_IDLE))
4226 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4227 * is driven by the tick):
4229 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4232 find_matching_se(&se, &pse);
4233 update_curr(cfs_rq_of(se));
4235 if (wakeup_preempt_entity(se, pse) == 1) {
4237 * Bias pick_next to pick the sched entity that is
4238 * triggering this preemption.
4240 if (!next_buddy_marked)
4241 set_next_buddy(pse);
4250 * Only set the backward buddy when the current task is still
4251 * on the rq. This can happen when a wakeup gets interleaved
4252 * with schedule on the ->pre_schedule() or idle_balance()
4253 * point, either of which can * drop the rq lock.
4255 * Also, during early boot the idle thread is in the fair class,
4256 * for obvious reasons its a bad idea to schedule back to it.
4258 if (unlikely(!se->on_rq || curr == rq->idle))
4261 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4265 static struct task_struct *pick_next_task_fair(struct rq *rq)
4267 struct task_struct *p;
4268 struct cfs_rq *cfs_rq = &rq->cfs;
4269 struct sched_entity *se;
4271 if (!cfs_rq->nr_running)
4275 se = pick_next_entity(cfs_rq);
4276 set_next_entity(cfs_rq, se);
4277 cfs_rq = group_cfs_rq(se);
4281 if (hrtick_enabled(rq))
4282 hrtick_start_fair(rq, p);
4288 * Account for a descheduled task:
4290 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4292 struct sched_entity *se = &prev->se;
4293 struct cfs_rq *cfs_rq;
4295 for_each_sched_entity(se) {
4296 cfs_rq = cfs_rq_of(se);
4297 put_prev_entity(cfs_rq, se);
4302 * sched_yield() is very simple
4304 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4306 static void yield_task_fair(struct rq *rq)
4308 struct task_struct *curr = rq->curr;
4309 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4310 struct sched_entity *se = &curr->se;
4313 * Are we the only task in the tree?
4315 if (unlikely(rq->nr_running == 1))
4318 clear_buddies(cfs_rq, se);
4320 if (curr->policy != SCHED_BATCH) {
4321 update_rq_clock(rq);
4323 * Update run-time statistics of the 'current'.
4325 update_curr(cfs_rq);
4327 * Tell update_rq_clock() that we've just updated,
4328 * so we don't do microscopic update in schedule()
4329 * and double the fastpath cost.
4331 rq->skip_clock_update = 1;
4337 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4339 struct sched_entity *se = &p->se;
4341 /* throttled hierarchies are not runnable */
4342 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4345 /* Tell the scheduler that we'd really like pse to run next. */
4348 yield_task_fair(rq);
4354 /**************************************************
4355 * Fair scheduling class load-balancing methods.
4359 * The purpose of load-balancing is to achieve the same basic fairness the
4360 * per-cpu scheduler provides, namely provide a proportional amount of compute
4361 * time to each task. This is expressed in the following equation:
4363 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4365 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4366 * W_i,0 is defined as:
4368 * W_i,0 = \Sum_j w_i,j (2)
4370 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4371 * is derived from the nice value as per prio_to_weight[].
4373 * The weight average is an exponential decay average of the instantaneous
4376 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4378 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4379 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4380 * can also include other factors [XXX].
4382 * To achieve this balance we define a measure of imbalance which follows
4383 * directly from (1):
4385 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4387 * We them move tasks around to minimize the imbalance. In the continuous
4388 * function space it is obvious this converges, in the discrete case we get
4389 * a few fun cases generally called infeasible weight scenarios.
4392 * - infeasible weights;
4393 * - local vs global optima in the discrete case. ]
4398 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4399 * for all i,j solution, we create a tree of cpus that follows the hardware
4400 * topology where each level pairs two lower groups (or better). This results
4401 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4402 * tree to only the first of the previous level and we decrease the frequency
4403 * of load-balance at each level inv. proportional to the number of cpus in
4409 * \Sum { --- * --- * 2^i } = O(n) (5)
4411 * `- size of each group
4412 * | | `- number of cpus doing load-balance
4414 * `- sum over all levels
4416 * Coupled with a limit on how many tasks we can migrate every balance pass,
4417 * this makes (5) the runtime complexity of the balancer.
4419 * An important property here is that each CPU is still (indirectly) connected
4420 * to every other cpu in at most O(log n) steps:
4422 * The adjacency matrix of the resulting graph is given by:
4425 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4428 * And you'll find that:
4430 * A^(log_2 n)_i,j != 0 for all i,j (7)
4432 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4433 * The task movement gives a factor of O(m), giving a convergence complexity
4436 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4441 * In order to avoid CPUs going idle while there's still work to do, new idle
4442 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4443 * tree itself instead of relying on other CPUs to bring it work.
4445 * This adds some complexity to both (5) and (8) but it reduces the total idle
4453 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4456 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4461 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4463 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4465 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4468 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4469 * rewrite all of this once again.]
4472 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4474 #define LBF_ALL_PINNED 0x01
4475 #define LBF_NEED_BREAK 0x02
4476 #define LBF_DST_PINNED 0x04
4477 #define LBF_SOME_PINNED 0x08
4480 struct sched_domain *sd;
4488 struct cpumask *dst_grpmask;
4490 enum cpu_idle_type idle;
4492 /* The set of CPUs under consideration for load-balancing */
4493 struct cpumask *cpus;
4498 unsigned int loop_break;
4499 unsigned int loop_max;
4503 * move_task - move a task from one runqueue to another runqueue.
4504 * Both runqueues must be locked.
4506 static void move_task(struct task_struct *p, struct lb_env *env)
4508 deactivate_task(env->src_rq, p, 0);
4509 set_task_cpu(p, env->dst_cpu);
4510 activate_task(env->dst_rq, p, 0);
4511 check_preempt_curr(env->dst_rq, p, 0);
4512 #ifdef CONFIG_NUMA_BALANCING
4513 if (p->numa_preferred_nid != -1) {
4514 int src_nid = cpu_to_node(env->src_cpu);
4515 int dst_nid = cpu_to_node(env->dst_cpu);
4518 * If the load balancer has moved the task then limit
4519 * migrations from taking place in the short term in
4520 * case this is a short-lived migration.
4522 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4523 p->numa_migrate_seq = 0;
4529 * Is this task likely cache-hot:
4532 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4536 if (p->sched_class != &fair_sched_class)
4539 if (unlikely(p->policy == SCHED_IDLE))
4543 * Buddy candidates are cache hot:
4545 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4546 (&p->se == cfs_rq_of(&p->se)->next ||
4547 &p->se == cfs_rq_of(&p->se)->last))
4550 if (sysctl_sched_migration_cost == -1)
4552 if (sysctl_sched_migration_cost == 0)
4555 delta = now - p->se.exec_start;
4557 return delta < (s64)sysctl_sched_migration_cost;
4560 #ifdef CONFIG_NUMA_BALANCING
4561 /* Returns true if the destination node has incurred more faults */
4562 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4564 int src_nid, dst_nid;
4566 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4567 !(env->sd->flags & SD_NUMA)) {
4571 src_nid = cpu_to_node(env->src_cpu);
4572 dst_nid = cpu_to_node(env->dst_cpu);
4574 if (src_nid == dst_nid ||
4575 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4578 if (dst_nid == p->numa_preferred_nid ||
4579 task_faults(p, dst_nid) > task_faults(p, src_nid))
4586 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4588 int src_nid, dst_nid;
4590 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4593 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4596 src_nid = cpu_to_node(env->src_cpu);
4597 dst_nid = cpu_to_node(env->dst_cpu);
4599 if (src_nid == dst_nid ||
4600 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4603 if (task_faults(p, dst_nid) < task_faults(p, src_nid))
4610 static inline bool migrate_improves_locality(struct task_struct *p,
4616 static inline bool migrate_degrades_locality(struct task_struct *p,
4624 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4627 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4629 int tsk_cache_hot = 0;
4631 * We do not migrate tasks that are:
4632 * 1) throttled_lb_pair, or
4633 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4634 * 3) running (obviously), or
4635 * 4) are cache-hot on their current CPU.
4637 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4640 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4643 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4645 env->flags |= LBF_SOME_PINNED;
4648 * Remember if this task can be migrated to any other cpu in
4649 * our sched_group. We may want to revisit it if we couldn't
4650 * meet load balance goals by pulling other tasks on src_cpu.
4652 * Also avoid computing new_dst_cpu if we have already computed
4653 * one in current iteration.
4655 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4658 /* Prevent to re-select dst_cpu via env's cpus */
4659 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4660 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4661 env->flags |= LBF_DST_PINNED;
4662 env->new_dst_cpu = cpu;
4670 /* Record that we found atleast one task that could run on dst_cpu */
4671 env->flags &= ~LBF_ALL_PINNED;
4673 if (task_running(env->src_rq, p)) {
4674 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4679 * Aggressive migration if:
4680 * 1) destination numa is preferred
4681 * 2) task is cache cold, or
4682 * 3) too many balance attempts have failed.
4684 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4686 tsk_cache_hot = migrate_degrades_locality(p, env);
4688 if (migrate_improves_locality(p, env)) {
4689 #ifdef CONFIG_SCHEDSTATS
4690 if (tsk_cache_hot) {
4691 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4692 schedstat_inc(p, se.statistics.nr_forced_migrations);
4698 if (!tsk_cache_hot ||
4699 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4701 if (tsk_cache_hot) {
4702 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4703 schedstat_inc(p, se.statistics.nr_forced_migrations);
4709 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4714 * move_one_task tries to move exactly one task from busiest to this_rq, as
4715 * part of active balancing operations within "domain".
4716 * Returns 1 if successful and 0 otherwise.
4718 * Called with both runqueues locked.
4720 static int move_one_task(struct lb_env *env)
4722 struct task_struct *p, *n;
4724 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4725 if (!can_migrate_task(p, env))
4730 * Right now, this is only the second place move_task()
4731 * is called, so we can safely collect move_task()
4732 * stats here rather than inside move_task().
4734 schedstat_inc(env->sd, lb_gained[env->idle]);
4740 static const unsigned int sched_nr_migrate_break = 32;
4743 * move_tasks tries to move up to imbalance weighted load from busiest to
4744 * this_rq, as part of a balancing operation within domain "sd".
4745 * Returns 1 if successful and 0 otherwise.
4747 * Called with both runqueues locked.
4749 static int move_tasks(struct lb_env *env)
4751 struct list_head *tasks = &env->src_rq->cfs_tasks;
4752 struct task_struct *p;
4756 if (env->imbalance <= 0)
4759 while (!list_empty(tasks)) {
4760 p = list_first_entry(tasks, struct task_struct, se.group_node);
4763 /* We've more or less seen every task there is, call it quits */
4764 if (env->loop > env->loop_max)
4767 /* take a breather every nr_migrate tasks */
4768 if (env->loop > env->loop_break) {
4769 env->loop_break += sched_nr_migrate_break;
4770 env->flags |= LBF_NEED_BREAK;
4774 if (!can_migrate_task(p, env))
4777 load = task_h_load(p);
4779 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4782 if ((load / 2) > env->imbalance)
4787 env->imbalance -= load;
4789 #ifdef CONFIG_PREEMPT
4791 * NEWIDLE balancing is a source of latency, so preemptible
4792 * kernels will stop after the first task is pulled to minimize
4793 * the critical section.
4795 if (env->idle == CPU_NEWLY_IDLE)
4800 * We only want to steal up to the prescribed amount of
4803 if (env->imbalance <= 0)
4808 list_move_tail(&p->se.group_node, tasks);
4812 * Right now, this is one of only two places move_task() is called,
4813 * so we can safely collect move_task() stats here rather than
4814 * inside move_task().
4816 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4821 #ifdef CONFIG_FAIR_GROUP_SCHED
4823 * update tg->load_weight by folding this cpu's load_avg
4825 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4827 struct sched_entity *se = tg->se[cpu];
4828 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4830 /* throttled entities do not contribute to load */
4831 if (throttled_hierarchy(cfs_rq))
4834 update_cfs_rq_blocked_load(cfs_rq, 1);
4837 update_entity_load_avg(se, 1);
4839 * We pivot on our runnable average having decayed to zero for
4840 * list removal. This generally implies that all our children
4841 * have also been removed (modulo rounding error or bandwidth
4842 * control); however, such cases are rare and we can fix these
4845 * TODO: fix up out-of-order children on enqueue.
4847 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4848 list_del_leaf_cfs_rq(cfs_rq);
4850 struct rq *rq = rq_of(cfs_rq);
4851 update_rq_runnable_avg(rq, rq->nr_running);
4855 static void update_blocked_averages(int cpu)
4857 struct rq *rq = cpu_rq(cpu);
4858 struct cfs_rq *cfs_rq;
4859 unsigned long flags;
4861 raw_spin_lock_irqsave(&rq->lock, flags);
4862 update_rq_clock(rq);
4864 * Iterates the task_group tree in a bottom up fashion, see
4865 * list_add_leaf_cfs_rq() for details.
4867 for_each_leaf_cfs_rq(rq, cfs_rq) {
4869 * Note: We may want to consider periodically releasing
4870 * rq->lock about these updates so that creating many task
4871 * groups does not result in continually extending hold time.
4873 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4876 raw_spin_unlock_irqrestore(&rq->lock, flags);
4880 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4881 * This needs to be done in a top-down fashion because the load of a child
4882 * group is a fraction of its parents load.
4884 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4886 struct rq *rq = rq_of(cfs_rq);
4887 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4888 unsigned long now = jiffies;
4891 if (cfs_rq->last_h_load_update == now)
4894 cfs_rq->h_load_next = NULL;
4895 for_each_sched_entity(se) {
4896 cfs_rq = cfs_rq_of(se);
4897 cfs_rq->h_load_next = se;
4898 if (cfs_rq->last_h_load_update == now)
4903 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4904 cfs_rq->last_h_load_update = now;
4907 while ((se = cfs_rq->h_load_next) != NULL) {
4908 load = cfs_rq->h_load;
4909 load = div64_ul(load * se->avg.load_avg_contrib,
4910 cfs_rq->runnable_load_avg + 1);
4911 cfs_rq = group_cfs_rq(se);
4912 cfs_rq->h_load = load;
4913 cfs_rq->last_h_load_update = now;
4917 static unsigned long task_h_load(struct task_struct *p)
4919 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4921 update_cfs_rq_h_load(cfs_rq);
4922 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4923 cfs_rq->runnable_load_avg + 1);
4926 static inline void update_blocked_averages(int cpu)
4930 static unsigned long task_h_load(struct task_struct *p)
4932 return p->se.avg.load_avg_contrib;
4936 /********** Helpers for find_busiest_group ************************/
4938 * sg_lb_stats - stats of a sched_group required for load_balancing
4940 struct sg_lb_stats {
4941 unsigned long avg_load; /*Avg load across the CPUs of the group */
4942 unsigned long group_load; /* Total load over the CPUs of the group */
4943 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4944 unsigned long load_per_task;
4945 unsigned long group_power;
4946 unsigned int sum_nr_running; /* Nr tasks running in the group */
4947 unsigned int group_capacity;
4948 unsigned int idle_cpus;
4949 unsigned int group_weight;
4950 int group_imb; /* Is there an imbalance in the group ? */
4951 int group_has_capacity; /* Is there extra capacity in the group? */
4955 * sd_lb_stats - Structure to store the statistics of a sched_domain
4956 * during load balancing.
4958 struct sd_lb_stats {
4959 struct sched_group *busiest; /* Busiest group in this sd */
4960 struct sched_group *local; /* Local group in this sd */
4961 unsigned long total_load; /* Total load of all groups in sd */
4962 unsigned long total_pwr; /* Total power of all groups in sd */
4963 unsigned long avg_load; /* Average load across all groups in sd */
4965 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4966 struct sg_lb_stats local_stat; /* Statistics of the local group */
4969 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4972 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4973 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4974 * We must however clear busiest_stat::avg_load because
4975 * update_sd_pick_busiest() reads this before assignment.
4977 *sds = (struct sd_lb_stats){
4989 * get_sd_load_idx - Obtain the load index for a given sched domain.
4990 * @sd: The sched_domain whose load_idx is to be obtained.
4991 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4993 * Return: The load index.
4995 static inline int get_sd_load_idx(struct sched_domain *sd,
4996 enum cpu_idle_type idle)
5002 load_idx = sd->busy_idx;
5005 case CPU_NEWLY_IDLE:
5006 load_idx = sd->newidle_idx;
5009 load_idx = sd->idle_idx;
5016 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5018 return SCHED_POWER_SCALE;
5021 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5023 return default_scale_freq_power(sd, cpu);
5026 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5028 unsigned long weight = sd->span_weight;
5029 unsigned long smt_gain = sd->smt_gain;
5036 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5038 return default_scale_smt_power(sd, cpu);
5041 static unsigned long scale_rt_power(int cpu)
5043 struct rq *rq = cpu_rq(cpu);
5044 u64 total, available, age_stamp, avg;
5047 * Since we're reading these variables without serialization make sure
5048 * we read them once before doing sanity checks on them.
5050 age_stamp = ACCESS_ONCE(rq->age_stamp);
5051 avg = ACCESS_ONCE(rq->rt_avg);
5053 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5055 if (unlikely(total < avg)) {
5056 /* Ensures that power won't end up being negative */
5059 available = total - avg;
5062 if (unlikely((s64)total < SCHED_POWER_SCALE))
5063 total = SCHED_POWER_SCALE;
5065 total >>= SCHED_POWER_SHIFT;
5067 return div_u64(available, total);
5070 static void update_cpu_power(struct sched_domain *sd, int cpu)
5072 unsigned long weight = sd->span_weight;
5073 unsigned long power = SCHED_POWER_SCALE;
5074 struct sched_group *sdg = sd->groups;
5076 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5077 if (sched_feat(ARCH_POWER))
5078 power *= arch_scale_smt_power(sd, cpu);
5080 power *= default_scale_smt_power(sd, cpu);
5082 power >>= SCHED_POWER_SHIFT;
5085 sdg->sgp->power_orig = power;
5087 if (sched_feat(ARCH_POWER))
5088 power *= arch_scale_freq_power(sd, cpu);
5090 power *= default_scale_freq_power(sd, cpu);
5092 power >>= SCHED_POWER_SHIFT;
5094 power *= scale_rt_power(cpu);
5095 power >>= SCHED_POWER_SHIFT;
5100 cpu_rq(cpu)->cpu_power = power;
5101 sdg->sgp->power = power;
5104 void update_group_power(struct sched_domain *sd, int cpu)
5106 struct sched_domain *child = sd->child;
5107 struct sched_group *group, *sdg = sd->groups;
5108 unsigned long power, power_orig;
5109 unsigned long interval;
5111 interval = msecs_to_jiffies(sd->balance_interval);
5112 interval = clamp(interval, 1UL, max_load_balance_interval);
5113 sdg->sgp->next_update = jiffies + interval;
5116 update_cpu_power(sd, cpu);
5120 power_orig = power = 0;
5122 if (child->flags & SD_OVERLAP) {
5124 * SD_OVERLAP domains cannot assume that child groups
5125 * span the current group.
5128 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5129 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5131 power_orig += sg->sgp->power_orig;
5132 power += sg->sgp->power;
5136 * !SD_OVERLAP domains can assume that child groups
5137 * span the current group.
5140 group = child->groups;
5142 power_orig += group->sgp->power_orig;
5143 power += group->sgp->power;
5144 group = group->next;
5145 } while (group != child->groups);
5148 sdg->sgp->power_orig = power_orig;
5149 sdg->sgp->power = power;
5153 * Try and fix up capacity for tiny siblings, this is needed when
5154 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5155 * which on its own isn't powerful enough.
5157 * See update_sd_pick_busiest() and check_asym_packing().
5160 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5163 * Only siblings can have significantly less than SCHED_POWER_SCALE
5165 if (!(sd->flags & SD_SHARE_CPUPOWER))
5169 * If ~90% of the cpu_power is still there, we're good.
5171 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5178 * Group imbalance indicates (and tries to solve) the problem where balancing
5179 * groups is inadequate due to tsk_cpus_allowed() constraints.
5181 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5182 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5185 * { 0 1 2 3 } { 4 5 6 7 }
5188 * If we were to balance group-wise we'd place two tasks in the first group and
5189 * two tasks in the second group. Clearly this is undesired as it will overload
5190 * cpu 3 and leave one of the cpus in the second group unused.
5192 * The current solution to this issue is detecting the skew in the first group
5193 * by noticing the lower domain failed to reach balance and had difficulty
5194 * moving tasks due to affinity constraints.
5196 * When this is so detected; this group becomes a candidate for busiest; see
5197 * update_sd_pick_busiest(). And calculcate_imbalance() and
5198 * find_busiest_group() avoid some of the usual balance conditions to allow it
5199 * to create an effective group imbalance.
5201 * This is a somewhat tricky proposition since the next run might not find the
5202 * group imbalance and decide the groups need to be balanced again. A most
5203 * subtle and fragile situation.
5206 static inline int sg_imbalanced(struct sched_group *group)
5208 return group->sgp->imbalance;
5212 * Compute the group capacity.
5214 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5215 * first dividing out the smt factor and computing the actual number of cores
5216 * and limit power unit capacity with that.
5218 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5220 unsigned int capacity, smt, cpus;
5221 unsigned int power, power_orig;
5223 power = group->sgp->power;
5224 power_orig = group->sgp->power_orig;
5225 cpus = group->group_weight;
5227 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5228 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5229 capacity = cpus / smt; /* cores */
5231 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5233 capacity = fix_small_capacity(env->sd, group);
5239 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5240 * @env: The load balancing environment.
5241 * @group: sched_group whose statistics are to be updated.
5242 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5243 * @local_group: Does group contain this_cpu.
5244 * @sgs: variable to hold the statistics for this group.
5246 static inline void update_sg_lb_stats(struct lb_env *env,
5247 struct sched_group *group, int load_idx,
5248 int local_group, struct sg_lb_stats *sgs)
5250 unsigned long nr_running;
5254 memset(sgs, 0, sizeof(*sgs));
5256 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5257 struct rq *rq = cpu_rq(i);
5259 nr_running = rq->nr_running;
5261 /* Bias balancing toward cpus of our domain */
5263 load = target_load(i, load_idx);
5265 load = source_load(i, load_idx);
5267 sgs->group_load += load;
5268 sgs->sum_nr_running += nr_running;
5269 sgs->sum_weighted_load += weighted_cpuload(i);
5274 /* Adjust by relative CPU power of the group */
5275 sgs->group_power = group->sgp->power;
5276 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5278 if (sgs->sum_nr_running)
5279 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5281 sgs->group_weight = group->group_weight;
5283 sgs->group_imb = sg_imbalanced(group);
5284 sgs->group_capacity = sg_capacity(env, group);
5286 if (sgs->group_capacity > sgs->sum_nr_running)
5287 sgs->group_has_capacity = 1;
5291 * update_sd_pick_busiest - return 1 on busiest group
5292 * @env: The load balancing environment.
5293 * @sds: sched_domain statistics
5294 * @sg: sched_group candidate to be checked for being the busiest
5295 * @sgs: sched_group statistics
5297 * Determine if @sg is a busier group than the previously selected
5300 * Return: %true if @sg is a busier group than the previously selected
5301 * busiest group. %false otherwise.
5303 static bool update_sd_pick_busiest(struct lb_env *env,
5304 struct sd_lb_stats *sds,
5305 struct sched_group *sg,
5306 struct sg_lb_stats *sgs)
5308 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5311 if (sgs->sum_nr_running > sgs->group_capacity)
5318 * ASYM_PACKING needs to move all the work to the lowest
5319 * numbered CPUs in the group, therefore mark all groups
5320 * higher than ourself as busy.
5322 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5323 env->dst_cpu < group_first_cpu(sg)) {
5327 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5335 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5336 * @env: The load balancing environment.
5337 * @balance: Should we balance.
5338 * @sds: variable to hold the statistics for this sched_domain.
5340 static inline void update_sd_lb_stats(struct lb_env *env,
5341 struct sd_lb_stats *sds)
5343 struct sched_domain *child = env->sd->child;
5344 struct sched_group *sg = env->sd->groups;
5345 struct sg_lb_stats tmp_sgs;
5346 int load_idx, prefer_sibling = 0;
5348 if (child && child->flags & SD_PREFER_SIBLING)
5351 load_idx = get_sd_load_idx(env->sd, env->idle);
5354 struct sg_lb_stats *sgs = &tmp_sgs;
5357 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5360 sgs = &sds->local_stat;
5362 if (env->idle != CPU_NEWLY_IDLE ||
5363 time_after_eq(jiffies, sg->sgp->next_update))
5364 update_group_power(env->sd, env->dst_cpu);
5367 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5373 * In case the child domain prefers tasks go to siblings
5374 * first, lower the sg capacity to one so that we'll try
5375 * and move all the excess tasks away. We lower the capacity
5376 * of a group only if the local group has the capacity to fit
5377 * these excess tasks, i.e. nr_running < group_capacity. The
5378 * extra check prevents the case where you always pull from the
5379 * heaviest group when it is already under-utilized (possible
5380 * with a large weight task outweighs the tasks on the system).
5382 if (prefer_sibling && sds->local &&
5383 sds->local_stat.group_has_capacity)
5384 sgs->group_capacity = min(sgs->group_capacity, 1U);
5386 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5388 sds->busiest_stat = *sgs;
5392 /* Now, start updating sd_lb_stats */
5393 sds->total_load += sgs->group_load;
5394 sds->total_pwr += sgs->group_power;
5397 } while (sg != env->sd->groups);
5401 * check_asym_packing - Check to see if the group is packed into the
5404 * This is primarily intended to used at the sibling level. Some
5405 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5406 * case of POWER7, it can move to lower SMT modes only when higher
5407 * threads are idle. When in lower SMT modes, the threads will
5408 * perform better since they share less core resources. Hence when we
5409 * have idle threads, we want them to be the higher ones.
5411 * This packing function is run on idle threads. It checks to see if
5412 * the busiest CPU in this domain (core in the P7 case) has a higher
5413 * CPU number than the packing function is being run on. Here we are
5414 * assuming lower CPU number will be equivalent to lower a SMT thread
5417 * Return: 1 when packing is required and a task should be moved to
5418 * this CPU. The amount of the imbalance is returned in *imbalance.
5420 * @env: The load balancing environment.
5421 * @sds: Statistics of the sched_domain which is to be packed
5423 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5427 if (!(env->sd->flags & SD_ASYM_PACKING))
5433 busiest_cpu = group_first_cpu(sds->busiest);
5434 if (env->dst_cpu > busiest_cpu)
5437 env->imbalance = DIV_ROUND_CLOSEST(
5438 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5445 * fix_small_imbalance - Calculate the minor imbalance that exists
5446 * amongst the groups of a sched_domain, during
5448 * @env: The load balancing environment.
5449 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5452 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5454 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5455 unsigned int imbn = 2;
5456 unsigned long scaled_busy_load_per_task;
5457 struct sg_lb_stats *local, *busiest;
5459 local = &sds->local_stat;
5460 busiest = &sds->busiest_stat;
5462 if (!local->sum_nr_running)
5463 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5464 else if (busiest->load_per_task > local->load_per_task)
5467 scaled_busy_load_per_task =
5468 (busiest->load_per_task * SCHED_POWER_SCALE) /
5469 busiest->group_power;
5471 if (busiest->avg_load + scaled_busy_load_per_task >=
5472 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5473 env->imbalance = busiest->load_per_task;
5478 * OK, we don't have enough imbalance to justify moving tasks,
5479 * however we may be able to increase total CPU power used by
5483 pwr_now += busiest->group_power *
5484 min(busiest->load_per_task, busiest->avg_load);
5485 pwr_now += local->group_power *
5486 min(local->load_per_task, local->avg_load);
5487 pwr_now /= SCHED_POWER_SCALE;
5489 /* Amount of load we'd subtract */
5490 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5491 busiest->group_power;
5492 if (busiest->avg_load > tmp) {
5493 pwr_move += busiest->group_power *
5494 min(busiest->load_per_task,
5495 busiest->avg_load - tmp);
5498 /* Amount of load we'd add */
5499 if (busiest->avg_load * busiest->group_power <
5500 busiest->load_per_task * SCHED_POWER_SCALE) {
5501 tmp = (busiest->avg_load * busiest->group_power) /
5504 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5507 pwr_move += local->group_power *
5508 min(local->load_per_task, local->avg_load + tmp);
5509 pwr_move /= SCHED_POWER_SCALE;
5511 /* Move if we gain throughput */
5512 if (pwr_move > pwr_now)
5513 env->imbalance = busiest->load_per_task;
5517 * calculate_imbalance - Calculate the amount of imbalance present within the
5518 * groups of a given sched_domain during load balance.
5519 * @env: load balance environment
5520 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5522 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5524 unsigned long max_pull, load_above_capacity = ~0UL;
5525 struct sg_lb_stats *local, *busiest;
5527 local = &sds->local_stat;
5528 busiest = &sds->busiest_stat;
5530 if (busiest->group_imb) {
5532 * In the group_imb case we cannot rely on group-wide averages
5533 * to ensure cpu-load equilibrium, look at wider averages. XXX
5535 busiest->load_per_task =
5536 min(busiest->load_per_task, sds->avg_load);
5540 * In the presence of smp nice balancing, certain scenarios can have
5541 * max load less than avg load(as we skip the groups at or below
5542 * its cpu_power, while calculating max_load..)
5544 if (busiest->avg_load <= sds->avg_load ||
5545 local->avg_load >= sds->avg_load) {
5547 return fix_small_imbalance(env, sds);
5550 if (!busiest->group_imb) {
5552 * Don't want to pull so many tasks that a group would go idle.
5553 * Except of course for the group_imb case, since then we might
5554 * have to drop below capacity to reach cpu-load equilibrium.
5556 load_above_capacity =
5557 (busiest->sum_nr_running - busiest->group_capacity);
5559 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5560 load_above_capacity /= busiest->group_power;
5564 * We're trying to get all the cpus to the average_load, so we don't
5565 * want to push ourselves above the average load, nor do we wish to
5566 * reduce the max loaded cpu below the average load. At the same time,
5567 * we also don't want to reduce the group load below the group capacity
5568 * (so that we can implement power-savings policies etc). Thus we look
5569 * for the minimum possible imbalance.
5571 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5573 /* How much load to actually move to equalise the imbalance */
5574 env->imbalance = min(
5575 max_pull * busiest->group_power,
5576 (sds->avg_load - local->avg_load) * local->group_power
5577 ) / SCHED_POWER_SCALE;
5580 * if *imbalance is less than the average load per runnable task
5581 * there is no guarantee that any tasks will be moved so we'll have
5582 * a think about bumping its value to force at least one task to be
5585 if (env->imbalance < busiest->load_per_task)
5586 return fix_small_imbalance(env, sds);
5589 /******* find_busiest_group() helpers end here *********************/
5592 * find_busiest_group - Returns the busiest group within the sched_domain
5593 * if there is an imbalance. If there isn't an imbalance, and
5594 * the user has opted for power-savings, it returns a group whose
5595 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5596 * such a group exists.
5598 * Also calculates the amount of weighted load which should be moved
5599 * to restore balance.
5601 * @env: The load balancing environment.
5603 * Return: - The busiest group if imbalance exists.
5604 * - If no imbalance and user has opted for power-savings balance,
5605 * return the least loaded group whose CPUs can be
5606 * put to idle by rebalancing its tasks onto our group.
5608 static struct sched_group *find_busiest_group(struct lb_env *env)
5610 struct sg_lb_stats *local, *busiest;
5611 struct sd_lb_stats sds;
5613 init_sd_lb_stats(&sds);
5616 * Compute the various statistics relavent for load balancing at
5619 update_sd_lb_stats(env, &sds);
5620 local = &sds.local_stat;
5621 busiest = &sds.busiest_stat;
5623 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5624 check_asym_packing(env, &sds))
5627 /* There is no busy sibling group to pull tasks from */
5628 if (!sds.busiest || busiest->sum_nr_running == 0)
5631 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5634 * If the busiest group is imbalanced the below checks don't
5635 * work because they assume all things are equal, which typically
5636 * isn't true due to cpus_allowed constraints and the like.
5638 if (busiest->group_imb)
5641 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5642 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5643 !busiest->group_has_capacity)
5647 * If the local group is more busy than the selected busiest group
5648 * don't try and pull any tasks.
5650 if (local->avg_load >= busiest->avg_load)
5654 * Don't pull any tasks if this group is already above the domain
5657 if (local->avg_load >= sds.avg_load)
5660 if (env->idle == CPU_IDLE) {
5662 * This cpu is idle. If the busiest group load doesn't
5663 * have more tasks than the number of available cpu's and
5664 * there is no imbalance between this and busiest group
5665 * wrt to idle cpu's, it is balanced.
5667 if ((local->idle_cpus < busiest->idle_cpus) &&
5668 busiest->sum_nr_running <= busiest->group_weight)
5672 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5673 * imbalance_pct to be conservative.
5675 if (100 * busiest->avg_load <=
5676 env->sd->imbalance_pct * local->avg_load)
5681 /* Looks like there is an imbalance. Compute it */
5682 calculate_imbalance(env, &sds);
5691 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5693 static struct rq *find_busiest_queue(struct lb_env *env,
5694 struct sched_group *group)
5696 struct rq *busiest = NULL, *rq;
5697 unsigned long busiest_load = 0, busiest_power = 1;
5700 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5701 unsigned long power = power_of(i);
5702 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5707 capacity = fix_small_capacity(env->sd, group);
5710 wl = weighted_cpuload(i);
5713 * When comparing with imbalance, use weighted_cpuload()
5714 * which is not scaled with the cpu power.
5716 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5720 * For the load comparisons with the other cpu's, consider
5721 * the weighted_cpuload() scaled with the cpu power, so that
5722 * the load can be moved away from the cpu that is potentially
5723 * running at a lower capacity.
5725 * Thus we're looking for max(wl_i / power_i), crosswise
5726 * multiplication to rid ourselves of the division works out
5727 * to: wl_i * power_j > wl_j * power_i; where j is our
5730 if (wl * busiest_power > busiest_load * power) {
5732 busiest_power = power;
5741 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5742 * so long as it is large enough.
5744 #define MAX_PINNED_INTERVAL 512
5746 /* Working cpumask for load_balance and load_balance_newidle. */
5747 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5749 static int need_active_balance(struct lb_env *env)
5751 struct sched_domain *sd = env->sd;
5753 if (env->idle == CPU_NEWLY_IDLE) {
5756 * ASYM_PACKING needs to force migrate tasks from busy but
5757 * higher numbered CPUs in order to pack all tasks in the
5758 * lowest numbered CPUs.
5760 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5764 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5767 static int active_load_balance_cpu_stop(void *data);
5769 static int should_we_balance(struct lb_env *env)
5771 struct sched_group *sg = env->sd->groups;
5772 struct cpumask *sg_cpus, *sg_mask;
5773 int cpu, balance_cpu = -1;
5776 * In the newly idle case, we will allow all the cpu's
5777 * to do the newly idle load balance.
5779 if (env->idle == CPU_NEWLY_IDLE)
5782 sg_cpus = sched_group_cpus(sg);
5783 sg_mask = sched_group_mask(sg);
5784 /* Try to find first idle cpu */
5785 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5786 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5793 if (balance_cpu == -1)
5794 balance_cpu = group_balance_cpu(sg);
5797 * First idle cpu or the first cpu(busiest) in this sched group
5798 * is eligible for doing load balancing at this and above domains.
5800 return balance_cpu == env->dst_cpu;
5804 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5805 * tasks if there is an imbalance.
5807 static int load_balance(int this_cpu, struct rq *this_rq,
5808 struct sched_domain *sd, enum cpu_idle_type idle,
5809 int *continue_balancing)
5811 int ld_moved, cur_ld_moved, active_balance = 0;
5812 struct sched_domain *sd_parent = sd->parent;
5813 struct sched_group *group;
5815 unsigned long flags;
5816 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5818 struct lb_env env = {
5820 .dst_cpu = this_cpu,
5822 .dst_grpmask = sched_group_cpus(sd->groups),
5824 .loop_break = sched_nr_migrate_break,
5829 * For NEWLY_IDLE load_balancing, we don't need to consider
5830 * other cpus in our group
5832 if (idle == CPU_NEWLY_IDLE)
5833 env.dst_grpmask = NULL;
5835 cpumask_copy(cpus, cpu_active_mask);
5837 schedstat_inc(sd, lb_count[idle]);
5840 if (!should_we_balance(&env)) {
5841 *continue_balancing = 0;
5845 group = find_busiest_group(&env);
5847 schedstat_inc(sd, lb_nobusyg[idle]);
5851 busiest = find_busiest_queue(&env, group);
5853 schedstat_inc(sd, lb_nobusyq[idle]);
5857 BUG_ON(busiest == env.dst_rq);
5859 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5862 if (busiest->nr_running > 1) {
5864 * Attempt to move tasks. If find_busiest_group has found
5865 * an imbalance but busiest->nr_running <= 1, the group is
5866 * still unbalanced. ld_moved simply stays zero, so it is
5867 * correctly treated as an imbalance.
5869 env.flags |= LBF_ALL_PINNED;
5870 env.src_cpu = busiest->cpu;
5871 env.src_rq = busiest;
5872 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5875 local_irq_save(flags);
5876 double_rq_lock(env.dst_rq, busiest);
5879 * cur_ld_moved - load moved in current iteration
5880 * ld_moved - cumulative load moved across iterations
5882 cur_ld_moved = move_tasks(&env);
5883 ld_moved += cur_ld_moved;
5884 double_rq_unlock(env.dst_rq, busiest);
5885 local_irq_restore(flags);
5888 * some other cpu did the load balance for us.
5890 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5891 resched_cpu(env.dst_cpu);
5893 if (env.flags & LBF_NEED_BREAK) {
5894 env.flags &= ~LBF_NEED_BREAK;
5899 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5900 * us and move them to an alternate dst_cpu in our sched_group
5901 * where they can run. The upper limit on how many times we
5902 * iterate on same src_cpu is dependent on number of cpus in our
5905 * This changes load balance semantics a bit on who can move
5906 * load to a given_cpu. In addition to the given_cpu itself
5907 * (or a ilb_cpu acting on its behalf where given_cpu is
5908 * nohz-idle), we now have balance_cpu in a position to move
5909 * load to given_cpu. In rare situations, this may cause
5910 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5911 * _independently_ and at _same_ time to move some load to
5912 * given_cpu) causing exceess load to be moved to given_cpu.
5913 * This however should not happen so much in practice and
5914 * moreover subsequent load balance cycles should correct the
5915 * excess load moved.
5917 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5919 /* Prevent to re-select dst_cpu via env's cpus */
5920 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5922 env.dst_rq = cpu_rq(env.new_dst_cpu);
5923 env.dst_cpu = env.new_dst_cpu;
5924 env.flags &= ~LBF_DST_PINNED;
5926 env.loop_break = sched_nr_migrate_break;
5929 * Go back to "more_balance" rather than "redo" since we
5930 * need to continue with same src_cpu.
5936 * We failed to reach balance because of affinity.
5939 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5941 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5942 *group_imbalance = 1;
5943 } else if (*group_imbalance)
5944 *group_imbalance = 0;
5947 /* All tasks on this runqueue were pinned by CPU affinity */
5948 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5949 cpumask_clear_cpu(cpu_of(busiest), cpus);
5950 if (!cpumask_empty(cpus)) {
5952 env.loop_break = sched_nr_migrate_break;
5960 schedstat_inc(sd, lb_failed[idle]);
5962 * Increment the failure counter only on periodic balance.
5963 * We do not want newidle balance, which can be very
5964 * frequent, pollute the failure counter causing
5965 * excessive cache_hot migrations and active balances.
5967 if (idle != CPU_NEWLY_IDLE)
5968 sd->nr_balance_failed++;
5970 if (need_active_balance(&env)) {
5971 raw_spin_lock_irqsave(&busiest->lock, flags);
5973 /* don't kick the active_load_balance_cpu_stop,
5974 * if the curr task on busiest cpu can't be
5977 if (!cpumask_test_cpu(this_cpu,
5978 tsk_cpus_allowed(busiest->curr))) {
5979 raw_spin_unlock_irqrestore(&busiest->lock,
5981 env.flags |= LBF_ALL_PINNED;
5982 goto out_one_pinned;
5986 * ->active_balance synchronizes accesses to
5987 * ->active_balance_work. Once set, it's cleared
5988 * only after active load balance is finished.
5990 if (!busiest->active_balance) {
5991 busiest->active_balance = 1;
5992 busiest->push_cpu = this_cpu;
5995 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5997 if (active_balance) {
5998 stop_one_cpu_nowait(cpu_of(busiest),
5999 active_load_balance_cpu_stop, busiest,
6000 &busiest->active_balance_work);
6004 * We've kicked active balancing, reset the failure
6007 sd->nr_balance_failed = sd->cache_nice_tries+1;
6010 sd->nr_balance_failed = 0;
6012 if (likely(!active_balance)) {
6013 /* We were unbalanced, so reset the balancing interval */
6014 sd->balance_interval = sd->min_interval;
6017 * If we've begun active balancing, start to back off. This
6018 * case may not be covered by the all_pinned logic if there
6019 * is only 1 task on the busy runqueue (because we don't call
6022 if (sd->balance_interval < sd->max_interval)
6023 sd->balance_interval *= 2;
6029 schedstat_inc(sd, lb_balanced[idle]);
6031 sd->nr_balance_failed = 0;
6034 /* tune up the balancing interval */
6035 if (((env.flags & LBF_ALL_PINNED) &&
6036 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6037 (sd->balance_interval < sd->max_interval))
6038 sd->balance_interval *= 2;
6046 * idle_balance is called by schedule() if this_cpu is about to become
6047 * idle. Attempts to pull tasks from other CPUs.
6049 void idle_balance(int this_cpu, struct rq *this_rq)
6051 struct sched_domain *sd;
6052 int pulled_task = 0;
6053 unsigned long next_balance = jiffies + HZ;
6056 this_rq->idle_stamp = rq_clock(this_rq);
6058 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6062 * Drop the rq->lock, but keep IRQ/preempt disabled.
6064 raw_spin_unlock(&this_rq->lock);
6066 update_blocked_averages(this_cpu);
6068 for_each_domain(this_cpu, sd) {
6069 unsigned long interval;
6070 int continue_balancing = 1;
6071 u64 t0, domain_cost;
6073 if (!(sd->flags & SD_LOAD_BALANCE))
6076 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6079 if (sd->flags & SD_BALANCE_NEWIDLE) {
6080 t0 = sched_clock_cpu(this_cpu);
6082 /* If we've pulled tasks over stop searching: */
6083 pulled_task = load_balance(this_cpu, this_rq,
6085 &continue_balancing);
6087 domain_cost = sched_clock_cpu(this_cpu) - t0;
6088 if (domain_cost > sd->max_newidle_lb_cost)
6089 sd->max_newidle_lb_cost = domain_cost;
6091 curr_cost += domain_cost;
6094 interval = msecs_to_jiffies(sd->balance_interval);
6095 if (time_after(next_balance, sd->last_balance + interval))
6096 next_balance = sd->last_balance + interval;
6098 this_rq->idle_stamp = 0;
6104 raw_spin_lock(&this_rq->lock);
6106 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6108 * We are going idle. next_balance may be set based on
6109 * a busy processor. So reset next_balance.
6111 this_rq->next_balance = next_balance;
6114 if (curr_cost > this_rq->max_idle_balance_cost)
6115 this_rq->max_idle_balance_cost = curr_cost;
6119 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6120 * running tasks off the busiest CPU onto idle CPUs. It requires at
6121 * least 1 task to be running on each physical CPU where possible, and
6122 * avoids physical / logical imbalances.
6124 static int active_load_balance_cpu_stop(void *data)
6126 struct rq *busiest_rq = data;
6127 int busiest_cpu = cpu_of(busiest_rq);
6128 int target_cpu = busiest_rq->push_cpu;
6129 struct rq *target_rq = cpu_rq(target_cpu);
6130 struct sched_domain *sd;
6132 raw_spin_lock_irq(&busiest_rq->lock);
6134 /* make sure the requested cpu hasn't gone down in the meantime */
6135 if (unlikely(busiest_cpu != smp_processor_id() ||
6136 !busiest_rq->active_balance))
6139 /* Is there any task to move? */
6140 if (busiest_rq->nr_running <= 1)
6144 * This condition is "impossible", if it occurs
6145 * we need to fix it. Originally reported by
6146 * Bjorn Helgaas on a 128-cpu setup.
6148 BUG_ON(busiest_rq == target_rq);
6150 /* move a task from busiest_rq to target_rq */
6151 double_lock_balance(busiest_rq, target_rq);
6153 /* Search for an sd spanning us and the target CPU. */
6155 for_each_domain(target_cpu, sd) {
6156 if ((sd->flags & SD_LOAD_BALANCE) &&
6157 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6162 struct lb_env env = {
6164 .dst_cpu = target_cpu,
6165 .dst_rq = target_rq,
6166 .src_cpu = busiest_rq->cpu,
6167 .src_rq = busiest_rq,
6171 schedstat_inc(sd, alb_count);
6173 if (move_one_task(&env))
6174 schedstat_inc(sd, alb_pushed);
6176 schedstat_inc(sd, alb_failed);
6179 double_unlock_balance(busiest_rq, target_rq);
6181 busiest_rq->active_balance = 0;
6182 raw_spin_unlock_irq(&busiest_rq->lock);
6186 #ifdef CONFIG_NO_HZ_COMMON
6188 * idle load balancing details
6189 * - When one of the busy CPUs notice that there may be an idle rebalancing
6190 * needed, they will kick the idle load balancer, which then does idle
6191 * load balancing for all the idle CPUs.
6194 cpumask_var_t idle_cpus_mask;
6196 unsigned long next_balance; /* in jiffy units */
6197 } nohz ____cacheline_aligned;
6199 static inline int find_new_ilb(int call_cpu)
6201 int ilb = cpumask_first(nohz.idle_cpus_mask);
6203 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6210 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6211 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6212 * CPU (if there is one).
6214 static void nohz_balancer_kick(int cpu)
6218 nohz.next_balance++;
6220 ilb_cpu = find_new_ilb(cpu);
6222 if (ilb_cpu >= nr_cpu_ids)
6225 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6228 * Use smp_send_reschedule() instead of resched_cpu().
6229 * This way we generate a sched IPI on the target cpu which
6230 * is idle. And the softirq performing nohz idle load balance
6231 * will be run before returning from the IPI.
6233 smp_send_reschedule(ilb_cpu);
6237 static inline void nohz_balance_exit_idle(int cpu)
6239 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6240 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6241 atomic_dec(&nohz.nr_cpus);
6242 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6246 static inline void set_cpu_sd_state_busy(void)
6248 struct sched_domain *sd;
6251 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6253 if (!sd || !sd->nohz_idle)
6257 for (; sd; sd = sd->parent)
6258 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6263 void set_cpu_sd_state_idle(void)
6265 struct sched_domain *sd;
6268 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6270 if (!sd || sd->nohz_idle)
6274 for (; sd; sd = sd->parent)
6275 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6281 * This routine will record that the cpu is going idle with tick stopped.
6282 * This info will be used in performing idle load balancing in the future.
6284 void nohz_balance_enter_idle(int cpu)
6287 * If this cpu is going down, then nothing needs to be done.
6289 if (!cpu_active(cpu))
6292 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6295 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6296 atomic_inc(&nohz.nr_cpus);
6297 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6300 static int sched_ilb_notifier(struct notifier_block *nfb,
6301 unsigned long action, void *hcpu)
6303 switch (action & ~CPU_TASKS_FROZEN) {
6305 nohz_balance_exit_idle(smp_processor_id());
6313 static DEFINE_SPINLOCK(balancing);
6316 * Scale the max load_balance interval with the number of CPUs in the system.
6317 * This trades load-balance latency on larger machines for less cross talk.
6319 void update_max_interval(void)
6321 max_load_balance_interval = HZ*num_online_cpus()/10;
6325 * It checks each scheduling domain to see if it is due to be balanced,
6326 * and initiates a balancing operation if so.
6328 * Balancing parameters are set up in init_sched_domains.
6330 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6332 int continue_balancing = 1;
6333 struct rq *rq = cpu_rq(cpu);
6334 unsigned long interval;
6335 struct sched_domain *sd;
6336 /* Earliest time when we have to do rebalance again */
6337 unsigned long next_balance = jiffies + 60*HZ;
6338 int update_next_balance = 0;
6339 int need_serialize, need_decay = 0;
6342 update_blocked_averages(cpu);
6345 for_each_domain(cpu, sd) {
6347 * Decay the newidle max times here because this is a regular
6348 * visit to all the domains. Decay ~1% per second.
6350 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6351 sd->max_newidle_lb_cost =
6352 (sd->max_newidle_lb_cost * 253) / 256;
6353 sd->next_decay_max_lb_cost = jiffies + HZ;
6356 max_cost += sd->max_newidle_lb_cost;
6358 if (!(sd->flags & SD_LOAD_BALANCE))
6362 * Stop the load balance at this level. There is another
6363 * CPU in our sched group which is doing load balancing more
6366 if (!continue_balancing) {
6372 interval = sd->balance_interval;
6373 if (idle != CPU_IDLE)
6374 interval *= sd->busy_factor;
6376 /* scale ms to jiffies */
6377 interval = msecs_to_jiffies(interval);
6378 interval = clamp(interval, 1UL, max_load_balance_interval);
6380 need_serialize = sd->flags & SD_SERIALIZE;
6382 if (need_serialize) {
6383 if (!spin_trylock(&balancing))
6387 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6388 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6390 * The LBF_DST_PINNED logic could have changed
6391 * env->dst_cpu, so we can't know our idle
6392 * state even if we migrated tasks. Update it.
6394 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6396 sd->last_balance = jiffies;
6399 spin_unlock(&balancing);
6401 if (time_after(next_balance, sd->last_balance + interval)) {
6402 next_balance = sd->last_balance + interval;
6403 update_next_balance = 1;
6408 * Ensure the rq-wide value also decays but keep it at a
6409 * reasonable floor to avoid funnies with rq->avg_idle.
6411 rq->max_idle_balance_cost =
6412 max((u64)sysctl_sched_migration_cost, max_cost);
6417 * next_balance will be updated only when there is a need.
6418 * When the cpu is attached to null domain for ex, it will not be
6421 if (likely(update_next_balance))
6422 rq->next_balance = next_balance;
6425 #ifdef CONFIG_NO_HZ_COMMON
6427 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6428 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6430 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6432 struct rq *this_rq = cpu_rq(this_cpu);
6436 if (idle != CPU_IDLE ||
6437 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6440 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6441 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6445 * If this cpu gets work to do, stop the load balancing
6446 * work being done for other cpus. Next load
6447 * balancing owner will pick it up.
6452 rq = cpu_rq(balance_cpu);
6454 raw_spin_lock_irq(&rq->lock);
6455 update_rq_clock(rq);
6456 update_idle_cpu_load(rq);
6457 raw_spin_unlock_irq(&rq->lock);
6459 rebalance_domains(balance_cpu, CPU_IDLE);
6461 if (time_after(this_rq->next_balance, rq->next_balance))
6462 this_rq->next_balance = rq->next_balance;
6464 nohz.next_balance = this_rq->next_balance;
6466 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6470 * Current heuristic for kicking the idle load balancer in the presence
6471 * of an idle cpu is the system.
6472 * - This rq has more than one task.
6473 * - At any scheduler domain level, this cpu's scheduler group has multiple
6474 * busy cpu's exceeding the group's power.
6475 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6476 * domain span are idle.
6478 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6480 unsigned long now = jiffies;
6481 struct sched_domain *sd;
6483 if (unlikely(idle_cpu(cpu)))
6487 * We may be recently in ticked or tickless idle mode. At the first
6488 * busy tick after returning from idle, we will update the busy stats.
6490 set_cpu_sd_state_busy();
6491 nohz_balance_exit_idle(cpu);
6494 * None are in tickless mode and hence no need for NOHZ idle load
6497 if (likely(!atomic_read(&nohz.nr_cpus)))
6500 if (time_before(now, nohz.next_balance))
6503 if (rq->nr_running >= 2)
6507 for_each_domain(cpu, sd) {
6508 struct sched_group *sg = sd->groups;
6509 struct sched_group_power *sgp = sg->sgp;
6510 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6512 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6513 goto need_kick_unlock;
6515 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6516 && (cpumask_first_and(nohz.idle_cpus_mask,
6517 sched_domain_span(sd)) < cpu))
6518 goto need_kick_unlock;
6520 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6532 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6536 * run_rebalance_domains is triggered when needed from the scheduler tick.
6537 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6539 static void run_rebalance_domains(struct softirq_action *h)
6541 int this_cpu = smp_processor_id();
6542 struct rq *this_rq = cpu_rq(this_cpu);
6543 enum cpu_idle_type idle = this_rq->idle_balance ?
6544 CPU_IDLE : CPU_NOT_IDLE;
6546 rebalance_domains(this_cpu, idle);
6549 * If this cpu has a pending nohz_balance_kick, then do the
6550 * balancing on behalf of the other idle cpus whose ticks are
6553 nohz_idle_balance(this_cpu, idle);
6556 static inline int on_null_domain(int cpu)
6558 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6562 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6564 void trigger_load_balance(struct rq *rq, int cpu)
6566 /* Don't need to rebalance while attached to NULL domain */
6567 if (time_after_eq(jiffies, rq->next_balance) &&
6568 likely(!on_null_domain(cpu)))
6569 raise_softirq(SCHED_SOFTIRQ);
6570 #ifdef CONFIG_NO_HZ_COMMON
6571 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6572 nohz_balancer_kick(cpu);
6576 static void rq_online_fair(struct rq *rq)
6581 static void rq_offline_fair(struct rq *rq)
6585 /* Ensure any throttled groups are reachable by pick_next_task */
6586 unthrottle_offline_cfs_rqs(rq);
6589 #endif /* CONFIG_SMP */
6592 * scheduler tick hitting a task of our scheduling class:
6594 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6596 struct cfs_rq *cfs_rq;
6597 struct sched_entity *se = &curr->se;
6599 for_each_sched_entity(se) {
6600 cfs_rq = cfs_rq_of(se);
6601 entity_tick(cfs_rq, se, queued);
6604 if (numabalancing_enabled)
6605 task_tick_numa(rq, curr);
6607 update_rq_runnable_avg(rq, 1);
6611 * called on fork with the child task as argument from the parent's context
6612 * - child not yet on the tasklist
6613 * - preemption disabled
6615 static void task_fork_fair(struct task_struct *p)
6617 struct cfs_rq *cfs_rq;
6618 struct sched_entity *se = &p->se, *curr;
6619 int this_cpu = smp_processor_id();
6620 struct rq *rq = this_rq();
6621 unsigned long flags;
6623 raw_spin_lock_irqsave(&rq->lock, flags);
6625 update_rq_clock(rq);
6627 cfs_rq = task_cfs_rq(current);
6628 curr = cfs_rq->curr;
6631 * Not only the cpu but also the task_group of the parent might have
6632 * been changed after parent->se.parent,cfs_rq were copied to
6633 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6634 * of child point to valid ones.
6637 __set_task_cpu(p, this_cpu);
6640 update_curr(cfs_rq);
6643 se->vruntime = curr->vruntime;
6644 place_entity(cfs_rq, se, 1);
6646 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6648 * Upon rescheduling, sched_class::put_prev_task() will place
6649 * 'current' within the tree based on its new key value.
6651 swap(curr->vruntime, se->vruntime);
6652 resched_task(rq->curr);
6655 se->vruntime -= cfs_rq->min_vruntime;
6657 raw_spin_unlock_irqrestore(&rq->lock, flags);
6661 * Priority of the task has changed. Check to see if we preempt
6665 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6671 * Reschedule if we are currently running on this runqueue and
6672 * our priority decreased, or if we are not currently running on
6673 * this runqueue and our priority is higher than the current's
6675 if (rq->curr == p) {
6676 if (p->prio > oldprio)
6677 resched_task(rq->curr);
6679 check_preempt_curr(rq, p, 0);
6682 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6684 struct sched_entity *se = &p->se;
6685 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6688 * Ensure the task's vruntime is normalized, so that when its
6689 * switched back to the fair class the enqueue_entity(.flags=0) will
6690 * do the right thing.
6692 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6693 * have normalized the vruntime, if it was !on_rq, then only when
6694 * the task is sleeping will it still have non-normalized vruntime.
6696 if (!se->on_rq && p->state != TASK_RUNNING) {
6698 * Fix up our vruntime so that the current sleep doesn't
6699 * cause 'unlimited' sleep bonus.
6701 place_entity(cfs_rq, se, 0);
6702 se->vruntime -= cfs_rq->min_vruntime;
6707 * Remove our load from contribution when we leave sched_fair
6708 * and ensure we don't carry in an old decay_count if we
6711 if (se->avg.decay_count) {
6712 __synchronize_entity_decay(se);
6713 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6719 * We switched to the sched_fair class.
6721 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6727 * We were most likely switched from sched_rt, so
6728 * kick off the schedule if running, otherwise just see
6729 * if we can still preempt the current task.
6732 resched_task(rq->curr);
6734 check_preempt_curr(rq, p, 0);
6737 /* Account for a task changing its policy or group.
6739 * This routine is mostly called to set cfs_rq->curr field when a task
6740 * migrates between groups/classes.
6742 static void set_curr_task_fair(struct rq *rq)
6744 struct sched_entity *se = &rq->curr->se;
6746 for_each_sched_entity(se) {
6747 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6749 set_next_entity(cfs_rq, se);
6750 /* ensure bandwidth has been allocated on our new cfs_rq */
6751 account_cfs_rq_runtime(cfs_rq, 0);
6755 void init_cfs_rq(struct cfs_rq *cfs_rq)
6757 cfs_rq->tasks_timeline = RB_ROOT;
6758 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6759 #ifndef CONFIG_64BIT
6760 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6763 atomic64_set(&cfs_rq->decay_counter, 1);
6764 atomic_long_set(&cfs_rq->removed_load, 0);
6768 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 static void task_move_group_fair(struct task_struct *p, int on_rq)
6771 struct cfs_rq *cfs_rq;
6773 * If the task was not on the rq at the time of this cgroup movement
6774 * it must have been asleep, sleeping tasks keep their ->vruntime
6775 * absolute on their old rq until wakeup (needed for the fair sleeper
6776 * bonus in place_entity()).
6778 * If it was on the rq, we've just 'preempted' it, which does convert
6779 * ->vruntime to a relative base.
6781 * Make sure both cases convert their relative position when migrating
6782 * to another cgroup's rq. This does somewhat interfere with the
6783 * fair sleeper stuff for the first placement, but who cares.
6786 * When !on_rq, vruntime of the task has usually NOT been normalized.
6787 * But there are some cases where it has already been normalized:
6789 * - Moving a forked child which is waiting for being woken up by
6790 * wake_up_new_task().
6791 * - Moving a task which has been woken up by try_to_wake_up() and
6792 * waiting for actually being woken up by sched_ttwu_pending().
6794 * To prevent boost or penalty in the new cfs_rq caused by delta
6795 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6797 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6801 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6802 set_task_rq(p, task_cpu(p));
6804 cfs_rq = cfs_rq_of(&p->se);
6805 p->se.vruntime += cfs_rq->min_vruntime;
6808 * migrate_task_rq_fair() will have removed our previous
6809 * contribution, but we must synchronize for ongoing future
6812 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6813 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6818 void free_fair_sched_group(struct task_group *tg)
6822 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6824 for_each_possible_cpu(i) {
6826 kfree(tg->cfs_rq[i]);
6835 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6837 struct cfs_rq *cfs_rq;
6838 struct sched_entity *se;
6841 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6844 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6848 tg->shares = NICE_0_LOAD;
6850 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6852 for_each_possible_cpu(i) {
6853 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6854 GFP_KERNEL, cpu_to_node(i));
6858 se = kzalloc_node(sizeof(struct sched_entity),
6859 GFP_KERNEL, cpu_to_node(i));
6863 init_cfs_rq(cfs_rq);
6864 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6875 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6877 struct rq *rq = cpu_rq(cpu);
6878 unsigned long flags;
6881 * Only empty task groups can be destroyed; so we can speculatively
6882 * check on_list without danger of it being re-added.
6884 if (!tg->cfs_rq[cpu]->on_list)
6887 raw_spin_lock_irqsave(&rq->lock, flags);
6888 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6889 raw_spin_unlock_irqrestore(&rq->lock, flags);
6892 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6893 struct sched_entity *se, int cpu,
6894 struct sched_entity *parent)
6896 struct rq *rq = cpu_rq(cpu);
6900 init_cfs_rq_runtime(cfs_rq);
6902 tg->cfs_rq[cpu] = cfs_rq;
6905 /* se could be NULL for root_task_group */
6910 se->cfs_rq = &rq->cfs;
6912 se->cfs_rq = parent->my_q;
6915 update_load_set(&se->load, 0);
6916 se->parent = parent;
6919 static DEFINE_MUTEX(shares_mutex);
6921 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6924 unsigned long flags;
6927 * We can't change the weight of the root cgroup.
6932 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6934 mutex_lock(&shares_mutex);
6935 if (tg->shares == shares)
6938 tg->shares = shares;
6939 for_each_possible_cpu(i) {
6940 struct rq *rq = cpu_rq(i);
6941 struct sched_entity *se;
6944 /* Propagate contribution to hierarchy */
6945 raw_spin_lock_irqsave(&rq->lock, flags);
6947 /* Possible calls to update_curr() need rq clock */
6948 update_rq_clock(rq);
6949 for_each_sched_entity(se)
6950 update_cfs_shares(group_cfs_rq(se));
6951 raw_spin_unlock_irqrestore(&rq->lock, flags);
6955 mutex_unlock(&shares_mutex);
6958 #else /* CONFIG_FAIR_GROUP_SCHED */
6960 void free_fair_sched_group(struct task_group *tg) { }
6962 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6967 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6969 #endif /* CONFIG_FAIR_GROUP_SCHED */
6972 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6974 struct sched_entity *se = &task->se;
6975 unsigned int rr_interval = 0;
6978 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6981 if (rq->cfs.load.weight)
6982 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6988 * All the scheduling class methods:
6990 const struct sched_class fair_sched_class = {
6991 .next = &idle_sched_class,
6992 .enqueue_task = enqueue_task_fair,
6993 .dequeue_task = dequeue_task_fair,
6994 .yield_task = yield_task_fair,
6995 .yield_to_task = yield_to_task_fair,
6997 .check_preempt_curr = check_preempt_wakeup,
6999 .pick_next_task = pick_next_task_fair,
7000 .put_prev_task = put_prev_task_fair,
7003 .select_task_rq = select_task_rq_fair,
7004 .migrate_task_rq = migrate_task_rq_fair,
7006 .rq_online = rq_online_fair,
7007 .rq_offline = rq_offline_fair,
7009 .task_waking = task_waking_fair,
7012 .set_curr_task = set_curr_task_fair,
7013 .task_tick = task_tick_fair,
7014 .task_fork = task_fork_fair,
7016 .prio_changed = prio_changed_fair,
7017 .switched_from = switched_from_fair,
7018 .switched_to = switched_to_fair,
7020 .get_rr_interval = get_rr_interval_fair,
7022 #ifdef CONFIG_FAIR_GROUP_SCHED
7023 .task_move_group = task_move_group_fair,
7027 #ifdef CONFIG_SCHED_DEBUG
7028 void print_cfs_stats(struct seq_file *m, int cpu)
7030 struct cfs_rq *cfs_rq;
7033 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7034 print_cfs_rq(m, cpu, cfs_rq);
7039 __init void init_sched_fair_class(void)
7042 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7044 #ifdef CONFIG_NO_HZ_COMMON
7045 nohz.next_balance = jiffies;
7046 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7047 cpu_notifier(sched_ilb_notifier, 0);