2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
829 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size = 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct *p)
839 unsigned long rss = 0;
840 unsigned long nr_scan_pages;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
848 rss = get_mm_rss(p->mm);
852 rss = round_up(rss, nr_scan_pages);
853 return rss / nr_scan_pages;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct *p)
861 unsigned int scan, floor;
862 unsigned int windows = 1;
864 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
865 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
866 floor = 1000 / windows;
868 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
869 return max_t(unsigned int, floor, scan);
872 static unsigned int task_scan_max(struct task_struct *p)
874 unsigned int smin = task_scan_min(p);
877 /* Watch for min being lower than max due to floor calculations */
878 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
879 return max(smin, smax);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
891 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
893 rq->nr_numa_running += (p->numa_preferred_nid != -1);
894 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
897 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
899 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
900 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
906 spinlock_t lock; /* nr_tasks, tasks */
909 struct list_head task_list;
912 atomic_long_t total_faults;
913 atomic_long_t faults[0];
916 pid_t task_numa_group_id(struct task_struct *p)
918 return p->numa_group ? p->numa_group->gid : 0;
921 static inline int task_faults_idx(int nid, int priv)
923 return 2 * nid + priv;
926 static inline unsigned long task_faults(struct task_struct *p, int nid)
931 return p->numa_faults[task_faults_idx(nid, 0)] +
932 p->numa_faults[task_faults_idx(nid, 1)];
935 static inline unsigned long group_faults(struct task_struct *p, int nid)
940 return atomic_long_read(&p->numa_group->faults[2*nid]) +
941 atomic_long_read(&p->numa_group->faults[2*nid+1]);
945 * These return the fraction of accesses done by a particular task, or
946 * task group, on a particular numa node. The group weight is given a
947 * larger multiplier, in order to group tasks together that are almost
948 * evenly spread out between numa nodes.
950 static inline unsigned long task_weight(struct task_struct *p, int nid)
952 unsigned long total_faults;
957 total_faults = p->total_numa_faults;
962 return 1000 * task_faults(p, nid) / total_faults;
965 static inline unsigned long group_weight(struct task_struct *p, int nid)
967 unsigned long total_faults;
972 total_faults = atomic_long_read(&p->numa_group->total_faults);
977 return 1000 * group_faults(p, nid) / total_faults;
980 static unsigned long weighted_cpuload(const int cpu);
981 static unsigned long source_load(int cpu, int type);
982 static unsigned long target_load(int cpu, int type);
983 static unsigned long power_of(int cpu);
984 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
986 /* Cached statistics for all CPUs within a node */
988 unsigned long nr_running;
991 /* Total compute capacity of CPUs on a node */
994 /* Approximate capacity in terms of runnable tasks on a node */
995 unsigned long capacity;
1000 * XXX borrowed from update_sg_lb_stats
1002 static void update_numa_stats(struct numa_stats *ns, int nid)
1006 memset(ns, 0, sizeof(*ns));
1007 for_each_cpu(cpu, cpumask_of_node(nid)) {
1008 struct rq *rq = cpu_rq(cpu);
1010 ns->nr_running += rq->nr_running;
1011 ns->load += weighted_cpuload(cpu);
1012 ns->power += power_of(cpu);
1015 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1016 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1017 ns->has_capacity = (ns->nr_running < ns->capacity);
1020 struct task_numa_env {
1021 struct task_struct *p;
1023 int src_cpu, src_nid;
1024 int dst_cpu, dst_nid;
1026 struct numa_stats src_stats, dst_stats;
1028 int imbalance_pct, idx;
1030 struct task_struct *best_task;
1035 static void task_numa_assign(struct task_numa_env *env,
1036 struct task_struct *p, long imp)
1039 put_task_struct(env->best_task);
1044 env->best_imp = imp;
1045 env->best_cpu = env->dst_cpu;
1049 * This checks if the overall compute and NUMA accesses of the system would
1050 * be improved if the source tasks was migrated to the target dst_cpu taking
1051 * into account that it might be best if task running on the dst_cpu should
1052 * be exchanged with the source task
1054 static void task_numa_compare(struct task_numa_env *env,
1055 long taskimp, long groupimp)
1057 struct rq *src_rq = cpu_rq(env->src_cpu);
1058 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1059 struct task_struct *cur;
1060 long dst_load, src_load;
1062 long imp = (groupimp > 0) ? groupimp : taskimp;
1065 cur = ACCESS_ONCE(dst_rq->curr);
1066 if (cur->pid == 0) /* idle */
1070 * "imp" is the fault differential for the source task between the
1071 * source and destination node. Calculate the total differential for
1072 * the source task and potential destination task. The more negative
1073 * the value is, the more rmeote accesses that would be expected to
1074 * be incurred if the tasks were swapped.
1077 /* Skip this swap candidate if cannot move to the source cpu */
1078 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1082 * If dst and source tasks are in the same NUMA group, or not
1083 * in any group then look only at task weights.
1085 if (cur->numa_group == env->p->numa_group) {
1086 imp = taskimp + task_weight(cur, env->src_nid) -
1087 task_weight(cur, env->dst_nid);
1089 * Add some hysteresis to prevent swapping the
1090 * tasks within a group over tiny differences.
1092 if (cur->numa_group)
1096 * Compare the group weights. If a task is all by
1097 * itself (not part of a group), use the task weight
1100 if (env->p->numa_group)
1105 if (cur->numa_group)
1106 imp += group_weight(cur, env->src_nid) -
1107 group_weight(cur, env->dst_nid);
1109 imp += task_weight(cur, env->src_nid) -
1110 task_weight(cur, env->dst_nid);
1114 if (imp < env->best_imp)
1118 /* Is there capacity at our destination? */
1119 if (env->src_stats.has_capacity &&
1120 !env->dst_stats.has_capacity)
1126 /* Balance doesn't matter much if we're running a task per cpu */
1127 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1131 * In the overloaded case, try and keep the load balanced.
1134 dst_load = env->dst_stats.load;
1135 src_load = env->src_stats.load;
1137 /* XXX missing power terms */
1138 load = task_h_load(env->p);
1143 load = task_h_load(cur);
1148 /* make src_load the smaller */
1149 if (dst_load < src_load)
1150 swap(dst_load, src_load);
1152 if (src_load * env->imbalance_pct < dst_load * 100)
1156 task_numa_assign(env, cur, imp);
1161 static void task_numa_find_cpu(struct task_numa_env *env,
1162 long taskimp, long groupimp)
1166 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1167 /* Skip this CPU if the source task cannot migrate */
1168 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1172 task_numa_compare(env, taskimp, groupimp);
1176 static int task_numa_migrate(struct task_struct *p)
1178 struct task_numa_env env = {
1181 .src_cpu = task_cpu(p),
1182 .src_nid = task_node(p),
1184 .imbalance_pct = 112,
1190 struct sched_domain *sd;
1191 unsigned long taskweight, groupweight;
1193 long taskimp, groupimp;
1196 * Pick the lowest SD_NUMA domain, as that would have the smallest
1197 * imbalance and would be the first to start moving tasks about.
1199 * And we want to avoid any moving of tasks about, as that would create
1200 * random movement of tasks -- counter the numa conditions we're trying
1204 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1205 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1208 taskweight = task_weight(p, env.src_nid);
1209 groupweight = group_weight(p, env.src_nid);
1210 update_numa_stats(&env.src_stats, env.src_nid);
1211 env.dst_nid = p->numa_preferred_nid;
1212 taskimp = task_weight(p, env.dst_nid) - taskweight;
1213 groupimp = group_weight(p, env.dst_nid) - groupweight;
1214 update_numa_stats(&env.dst_stats, env.dst_nid);
1216 /* If the preferred nid has capacity, try to use it. */
1217 if (env.dst_stats.has_capacity)
1218 task_numa_find_cpu(&env, taskimp, groupimp);
1220 /* No space available on the preferred nid. Look elsewhere. */
1221 if (env.best_cpu == -1) {
1222 for_each_online_node(nid) {
1223 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1226 /* Only consider nodes where both task and groups benefit */
1227 taskimp = task_weight(p, nid) - taskweight;
1228 groupimp = group_weight(p, nid) - groupweight;
1229 if (taskimp < 0 && groupimp < 0)
1233 update_numa_stats(&env.dst_stats, env.dst_nid);
1234 task_numa_find_cpu(&env, taskimp, groupimp);
1238 /* No better CPU than the current one was found. */
1239 if (env.best_cpu == -1)
1242 sched_setnuma(p, env.dst_nid);
1244 if (env.best_task == NULL) {
1245 int ret = migrate_task_to(p, env.best_cpu);
1249 ret = migrate_swap(p, env.best_task);
1250 put_task_struct(env.best_task);
1254 /* Attempt to migrate a task to a CPU on the preferred node. */
1255 static void numa_migrate_preferred(struct task_struct *p)
1257 /* Success if task is already running on preferred CPU */
1258 p->numa_migrate_retry = 0;
1259 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1261 * If migration is temporarily disabled due to a task migration
1262 * then re-enable it now as the task is running on its
1263 * preferred node and memory should migrate locally
1265 if (!p->numa_migrate_seq)
1266 p->numa_migrate_seq++;
1270 /* This task has no NUMA fault statistics yet */
1271 if (unlikely(p->numa_preferred_nid == -1))
1274 /* Otherwise, try migrate to a CPU on the preferred node */
1275 if (task_numa_migrate(p) != 0)
1276 p->numa_migrate_retry = jiffies + HZ*5;
1279 static void task_numa_placement(struct task_struct *p)
1281 int seq, nid, max_nid = -1, max_group_nid = -1;
1282 unsigned long max_faults = 0, max_group_faults = 0;
1283 spinlock_t *group_lock = NULL;
1285 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1286 if (p->numa_scan_seq == seq)
1288 p->numa_scan_seq = seq;
1289 p->numa_migrate_seq++;
1290 p->numa_scan_period_max = task_scan_max(p);
1292 /* If the task is part of a group prevent parallel updates to group stats */
1293 if (p->numa_group) {
1294 group_lock = &p->numa_group->lock;
1295 spin_lock(group_lock);
1298 /* Find the node with the highest number of faults */
1299 for_each_online_node(nid) {
1300 unsigned long faults = 0, group_faults = 0;
1303 for (priv = 0; priv < 2; priv++) {
1306 i = task_faults_idx(nid, priv);
1307 diff = -p->numa_faults[i];
1309 /* Decay existing window, copy faults since last scan */
1310 p->numa_faults[i] >>= 1;
1311 p->numa_faults[i] += p->numa_faults_buffer[i];
1312 p->numa_faults_buffer[i] = 0;
1314 faults += p->numa_faults[i];
1315 diff += p->numa_faults[i];
1316 p->total_numa_faults += diff;
1317 if (p->numa_group) {
1318 /* safe because we can only change our own group */
1319 atomic_long_add(diff, &p->numa_group->faults[i]);
1320 atomic_long_add(diff, &p->numa_group->total_faults);
1321 group_faults += atomic_long_read(&p->numa_group->faults[i]);
1325 if (faults > max_faults) {
1326 max_faults = faults;
1330 if (group_faults > max_group_faults) {
1331 max_group_faults = group_faults;
1332 max_group_nid = nid;
1336 if (p->numa_group) {
1338 * If the preferred task and group nids are different,
1339 * iterate over the nodes again to find the best place.
1341 if (max_nid != max_group_nid) {
1342 unsigned long weight, max_weight = 0;
1344 for_each_online_node(nid) {
1345 weight = task_weight(p, nid) + group_weight(p, nid);
1346 if (weight > max_weight) {
1347 max_weight = weight;
1353 spin_unlock(group_lock);
1356 /* Preferred node as the node with the most faults */
1357 if (max_faults && max_nid != p->numa_preferred_nid) {
1358 /* Update the preferred nid and migrate task if possible */
1359 sched_setnuma(p, max_nid);
1360 numa_migrate_preferred(p);
1364 static inline int get_numa_group(struct numa_group *grp)
1366 return atomic_inc_not_zero(&grp->refcount);
1369 static inline void put_numa_group(struct numa_group *grp)
1371 if (atomic_dec_and_test(&grp->refcount))
1372 kfree_rcu(grp, rcu);
1375 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1381 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1384 static void task_numa_group(struct task_struct *p, int cpupid, int flags)
1386 struct numa_group *grp, *my_grp;
1387 struct task_struct *tsk;
1389 int cpu = cpupid_to_cpu(cpupid);
1392 if (unlikely(!p->numa_group)) {
1393 unsigned int size = sizeof(struct numa_group) +
1394 2*nr_node_ids*sizeof(atomic_long_t);
1396 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1400 atomic_set(&grp->refcount, 1);
1401 spin_lock_init(&grp->lock);
1402 INIT_LIST_HEAD(&grp->task_list);
1405 for (i = 0; i < 2*nr_node_ids; i++)
1406 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1408 atomic_long_set(&grp->total_faults, p->total_numa_faults);
1410 list_add(&p->numa_entry, &grp->task_list);
1412 rcu_assign_pointer(p->numa_group, grp);
1416 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1418 if (!cpupid_match_pid(tsk, cpupid))
1421 grp = rcu_dereference(tsk->numa_group);
1425 my_grp = p->numa_group;
1430 * Only join the other group if its bigger; if we're the bigger group,
1431 * the other task will join us.
1433 if (my_grp->nr_tasks > grp->nr_tasks)
1437 * Tie-break on the grp address.
1439 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1442 /* Always join threads in the same process. */
1443 if (tsk->mm == current->mm)
1446 /* Simple filter to avoid false positives due to PID collisions */
1447 if (flags & TNF_SHARED)
1450 if (join && !get_numa_group(grp))
1459 for (i = 0; i < 2*nr_node_ids; i++) {
1460 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1461 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1463 atomic_long_sub(p->total_numa_faults, &my_grp->total_faults);
1464 atomic_long_add(p->total_numa_faults, &grp->total_faults);
1466 double_lock(&my_grp->lock, &grp->lock);
1468 list_move(&p->numa_entry, &grp->task_list);
1472 spin_unlock(&my_grp->lock);
1473 spin_unlock(&grp->lock);
1475 rcu_assign_pointer(p->numa_group, grp);
1477 put_numa_group(my_grp);
1480 void task_numa_free(struct task_struct *p)
1482 struct numa_group *grp = p->numa_group;
1484 void *numa_faults = p->numa_faults;
1487 for (i = 0; i < 2*nr_node_ids; i++)
1488 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1490 atomic_long_sub(p->total_numa_faults, &grp->total_faults);
1492 spin_lock(&grp->lock);
1493 list_del(&p->numa_entry);
1495 spin_unlock(&grp->lock);
1496 rcu_assign_pointer(p->numa_group, NULL);
1497 put_numa_group(grp);
1500 p->numa_faults = NULL;
1501 p->numa_faults_buffer = NULL;
1506 * Got a PROT_NONE fault for a page on @node.
1508 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1510 struct task_struct *p = current;
1511 bool migrated = flags & TNF_MIGRATED;
1514 if (!numabalancing_enabled)
1517 /* for example, ksmd faulting in a user's mm */
1521 /* Do not worry about placement if exiting */
1522 if (p->state == TASK_DEAD)
1525 /* Allocate buffer to track faults on a per-node basis */
1526 if (unlikely(!p->numa_faults)) {
1527 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1529 /* numa_faults and numa_faults_buffer share the allocation */
1530 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1531 if (!p->numa_faults)
1534 BUG_ON(p->numa_faults_buffer);
1535 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1536 p->total_numa_faults = 0;
1540 * First accesses are treated as private, otherwise consider accesses
1541 * to be private if the accessing pid has not changed
1543 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1546 priv = cpupid_match_pid(p, last_cpupid);
1547 if (!priv && !(flags & TNF_NO_GROUP))
1548 task_numa_group(p, last_cpupid, flags);
1552 * If pages are properly placed (did not migrate) then scan slower.
1553 * This is reset periodically in case of phase changes
1556 /* Initialise if necessary */
1557 if (!p->numa_scan_period_max)
1558 p->numa_scan_period_max = task_scan_max(p);
1560 p->numa_scan_period = min(p->numa_scan_period_max,
1561 p->numa_scan_period + 10);
1564 task_numa_placement(p);
1566 /* Retry task to preferred node migration if it previously failed */
1567 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1568 numa_migrate_preferred(p);
1571 p->numa_pages_migrated += pages;
1573 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1576 static void reset_ptenuma_scan(struct task_struct *p)
1578 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1579 p->mm->numa_scan_offset = 0;
1583 * The expensive part of numa migration is done from task_work context.
1584 * Triggered from task_tick_numa().
1586 void task_numa_work(struct callback_head *work)
1588 unsigned long migrate, next_scan, now = jiffies;
1589 struct task_struct *p = current;
1590 struct mm_struct *mm = p->mm;
1591 struct vm_area_struct *vma;
1592 unsigned long start, end;
1593 unsigned long nr_pte_updates = 0;
1596 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1598 work->next = work; /* protect against double add */
1600 * Who cares about NUMA placement when they're dying.
1602 * NOTE: make sure not to dereference p->mm before this check,
1603 * exit_task_work() happens _after_ exit_mm() so we could be called
1604 * without p->mm even though we still had it when we enqueued this
1607 if (p->flags & PF_EXITING)
1610 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1611 mm->numa_next_scan = now +
1612 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1613 mm->numa_next_reset = now +
1614 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1618 * Reset the scan period if enough time has gone by. Objective is that
1619 * scanning will be reduced if pages are properly placed. As tasks
1620 * can enter different phases this needs to be re-examined. Lacking
1621 * proper tracking of reference behaviour, this blunt hammer is used.
1623 migrate = mm->numa_next_reset;
1624 if (time_after(now, migrate)) {
1625 p->numa_scan_period = task_scan_min(p);
1626 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1627 xchg(&mm->numa_next_reset, next_scan);
1631 * Enforce maximal scan/migration frequency..
1633 migrate = mm->numa_next_scan;
1634 if (time_before(now, migrate))
1637 if (p->numa_scan_period == 0) {
1638 p->numa_scan_period_max = task_scan_max(p);
1639 p->numa_scan_period = task_scan_min(p);
1642 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1643 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1647 * Delay this task enough that another task of this mm will likely win
1648 * the next time around.
1650 p->node_stamp += 2 * TICK_NSEC;
1652 start = mm->numa_scan_offset;
1653 pages = sysctl_numa_balancing_scan_size;
1654 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1658 down_read(&mm->mmap_sem);
1659 vma = find_vma(mm, start);
1661 reset_ptenuma_scan(p);
1665 for (; vma; vma = vma->vm_next) {
1666 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1670 * Shared library pages mapped by multiple processes are not
1671 * migrated as it is expected they are cache replicated. Avoid
1672 * hinting faults in read-only file-backed mappings or the vdso
1673 * as migrating the pages will be of marginal benefit.
1676 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1680 start = max(start, vma->vm_start);
1681 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1682 end = min(end, vma->vm_end);
1683 nr_pte_updates += change_prot_numa(vma, start, end);
1686 * Scan sysctl_numa_balancing_scan_size but ensure that
1687 * at least one PTE is updated so that unused virtual
1688 * address space is quickly skipped.
1691 pages -= (end - start) >> PAGE_SHIFT;
1696 } while (end != vma->vm_end);
1701 * If the whole process was scanned without updates then no NUMA
1702 * hinting faults are being recorded and scan rate should be lower.
1704 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1705 p->numa_scan_period = min(p->numa_scan_period_max,
1706 p->numa_scan_period << 1);
1708 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1709 mm->numa_next_scan = next_scan;
1713 * It is possible to reach the end of the VMA list but the last few
1714 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1715 * would find the !migratable VMA on the next scan but not reset the
1716 * scanner to the start so check it now.
1719 mm->numa_scan_offset = start;
1721 reset_ptenuma_scan(p);
1722 up_read(&mm->mmap_sem);
1726 * Drive the periodic memory faults..
1728 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1730 struct callback_head *work = &curr->numa_work;
1734 * We don't care about NUMA placement if we don't have memory.
1736 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1740 * Using runtime rather than walltime has the dual advantage that
1741 * we (mostly) drive the selection from busy threads and that the
1742 * task needs to have done some actual work before we bother with
1745 now = curr->se.sum_exec_runtime;
1746 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1748 if (now - curr->node_stamp > period) {
1749 if (!curr->node_stamp)
1750 curr->numa_scan_period = task_scan_min(curr);
1751 curr->node_stamp += period;
1753 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1754 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1755 task_work_add(curr, work, true);
1760 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1764 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1768 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1771 #endif /* CONFIG_NUMA_BALANCING */
1774 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1776 update_load_add(&cfs_rq->load, se->load.weight);
1777 if (!parent_entity(se))
1778 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1780 if (entity_is_task(se)) {
1781 struct rq *rq = rq_of(cfs_rq);
1783 account_numa_enqueue(rq, task_of(se));
1784 list_add(&se->group_node, &rq->cfs_tasks);
1787 cfs_rq->nr_running++;
1791 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1793 update_load_sub(&cfs_rq->load, se->load.weight);
1794 if (!parent_entity(se))
1795 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1796 if (entity_is_task(se)) {
1797 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1798 list_del_init(&se->group_node);
1800 cfs_rq->nr_running--;
1803 #ifdef CONFIG_FAIR_GROUP_SCHED
1805 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1810 * Use this CPU's actual weight instead of the last load_contribution
1811 * to gain a more accurate current total weight. See
1812 * update_cfs_rq_load_contribution().
1814 tg_weight = atomic_long_read(&tg->load_avg);
1815 tg_weight -= cfs_rq->tg_load_contrib;
1816 tg_weight += cfs_rq->load.weight;
1821 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1823 long tg_weight, load, shares;
1825 tg_weight = calc_tg_weight(tg, cfs_rq);
1826 load = cfs_rq->load.weight;
1828 shares = (tg->shares * load);
1830 shares /= tg_weight;
1832 if (shares < MIN_SHARES)
1833 shares = MIN_SHARES;
1834 if (shares > tg->shares)
1835 shares = tg->shares;
1839 # else /* CONFIG_SMP */
1840 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1844 # endif /* CONFIG_SMP */
1845 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1846 unsigned long weight)
1849 /* commit outstanding execution time */
1850 if (cfs_rq->curr == se)
1851 update_curr(cfs_rq);
1852 account_entity_dequeue(cfs_rq, se);
1855 update_load_set(&se->load, weight);
1858 account_entity_enqueue(cfs_rq, se);
1861 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1863 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1865 struct task_group *tg;
1866 struct sched_entity *se;
1870 se = tg->se[cpu_of(rq_of(cfs_rq))];
1871 if (!se || throttled_hierarchy(cfs_rq))
1874 if (likely(se->load.weight == tg->shares))
1877 shares = calc_cfs_shares(cfs_rq, tg);
1879 reweight_entity(cfs_rq_of(se), se, shares);
1881 #else /* CONFIG_FAIR_GROUP_SCHED */
1882 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1885 #endif /* CONFIG_FAIR_GROUP_SCHED */
1889 * We choose a half-life close to 1 scheduling period.
1890 * Note: The tables below are dependent on this value.
1892 #define LOAD_AVG_PERIOD 32
1893 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1894 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1896 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1897 static const u32 runnable_avg_yN_inv[] = {
1898 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1899 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1900 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1901 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1902 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1903 0x85aac367, 0x82cd8698,
1907 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1908 * over-estimates when re-combining.
1910 static const u32 runnable_avg_yN_sum[] = {
1911 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1912 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1913 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1918 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1920 static __always_inline u64 decay_load(u64 val, u64 n)
1922 unsigned int local_n;
1926 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1929 /* after bounds checking we can collapse to 32-bit */
1933 * As y^PERIOD = 1/2, we can combine
1934 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1935 * With a look-up table which covers k^n (n<PERIOD)
1937 * To achieve constant time decay_load.
1939 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1940 val >>= local_n / LOAD_AVG_PERIOD;
1941 local_n %= LOAD_AVG_PERIOD;
1944 val *= runnable_avg_yN_inv[local_n];
1945 /* We don't use SRR here since we always want to round down. */
1950 * For updates fully spanning n periods, the contribution to runnable
1951 * average will be: \Sum 1024*y^n
1953 * We can compute this reasonably efficiently by combining:
1954 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1956 static u32 __compute_runnable_contrib(u64 n)
1960 if (likely(n <= LOAD_AVG_PERIOD))
1961 return runnable_avg_yN_sum[n];
1962 else if (unlikely(n >= LOAD_AVG_MAX_N))
1963 return LOAD_AVG_MAX;
1965 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1967 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1968 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1970 n -= LOAD_AVG_PERIOD;
1971 } while (n > LOAD_AVG_PERIOD);
1973 contrib = decay_load(contrib, n);
1974 return contrib + runnable_avg_yN_sum[n];
1978 * We can represent the historical contribution to runnable average as the
1979 * coefficients of a geometric series. To do this we sub-divide our runnable
1980 * history into segments of approximately 1ms (1024us); label the segment that
1981 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1983 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1985 * (now) (~1ms ago) (~2ms ago)
1987 * Let u_i denote the fraction of p_i that the entity was runnable.
1989 * We then designate the fractions u_i as our co-efficients, yielding the
1990 * following representation of historical load:
1991 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1993 * We choose y based on the with of a reasonably scheduling period, fixing:
1996 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1997 * approximately half as much as the contribution to load within the last ms
2000 * When a period "rolls over" and we have new u_0`, multiplying the previous
2001 * sum again by y is sufficient to update:
2002 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2003 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2005 static __always_inline int __update_entity_runnable_avg(u64 now,
2006 struct sched_avg *sa,
2010 u32 runnable_contrib;
2011 int delta_w, decayed = 0;
2013 delta = now - sa->last_runnable_update;
2015 * This should only happen when time goes backwards, which it
2016 * unfortunately does during sched clock init when we swap over to TSC.
2018 if ((s64)delta < 0) {
2019 sa->last_runnable_update = now;
2024 * Use 1024ns as the unit of measurement since it's a reasonable
2025 * approximation of 1us and fast to compute.
2030 sa->last_runnable_update = now;
2032 /* delta_w is the amount already accumulated against our next period */
2033 delta_w = sa->runnable_avg_period % 1024;
2034 if (delta + delta_w >= 1024) {
2035 /* period roll-over */
2039 * Now that we know we're crossing a period boundary, figure
2040 * out how much from delta we need to complete the current
2041 * period and accrue it.
2043 delta_w = 1024 - delta_w;
2045 sa->runnable_avg_sum += delta_w;
2046 sa->runnable_avg_period += delta_w;
2050 /* Figure out how many additional periods this update spans */
2051 periods = delta / 1024;
2054 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2056 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2059 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2060 runnable_contrib = __compute_runnable_contrib(periods);
2062 sa->runnable_avg_sum += runnable_contrib;
2063 sa->runnable_avg_period += runnable_contrib;
2066 /* Remainder of delta accrued against u_0` */
2068 sa->runnable_avg_sum += delta;
2069 sa->runnable_avg_period += delta;
2074 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2075 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2077 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2078 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2080 decays -= se->avg.decay_count;
2084 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2085 se->avg.decay_count = 0;
2090 #ifdef CONFIG_FAIR_GROUP_SCHED
2091 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2094 struct task_group *tg = cfs_rq->tg;
2097 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2098 tg_contrib -= cfs_rq->tg_load_contrib;
2100 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2101 atomic_long_add(tg_contrib, &tg->load_avg);
2102 cfs_rq->tg_load_contrib += tg_contrib;
2107 * Aggregate cfs_rq runnable averages into an equivalent task_group
2108 * representation for computing load contributions.
2110 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2111 struct cfs_rq *cfs_rq)
2113 struct task_group *tg = cfs_rq->tg;
2116 /* The fraction of a cpu used by this cfs_rq */
2117 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2118 sa->runnable_avg_period + 1);
2119 contrib -= cfs_rq->tg_runnable_contrib;
2121 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2122 atomic_add(contrib, &tg->runnable_avg);
2123 cfs_rq->tg_runnable_contrib += contrib;
2127 static inline void __update_group_entity_contrib(struct sched_entity *se)
2129 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2130 struct task_group *tg = cfs_rq->tg;
2135 contrib = cfs_rq->tg_load_contrib * tg->shares;
2136 se->avg.load_avg_contrib = div_u64(contrib,
2137 atomic_long_read(&tg->load_avg) + 1);
2140 * For group entities we need to compute a correction term in the case
2141 * that they are consuming <1 cpu so that we would contribute the same
2142 * load as a task of equal weight.
2144 * Explicitly co-ordinating this measurement would be expensive, but
2145 * fortunately the sum of each cpus contribution forms a usable
2146 * lower-bound on the true value.
2148 * Consider the aggregate of 2 contributions. Either they are disjoint
2149 * (and the sum represents true value) or they are disjoint and we are
2150 * understating by the aggregate of their overlap.
2152 * Extending this to N cpus, for a given overlap, the maximum amount we
2153 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2154 * cpus that overlap for this interval and w_i is the interval width.
2156 * On a small machine; the first term is well-bounded which bounds the
2157 * total error since w_i is a subset of the period. Whereas on a
2158 * larger machine, while this first term can be larger, if w_i is the
2159 * of consequential size guaranteed to see n_i*w_i quickly converge to
2160 * our upper bound of 1-cpu.
2162 runnable_avg = atomic_read(&tg->runnable_avg);
2163 if (runnable_avg < NICE_0_LOAD) {
2164 se->avg.load_avg_contrib *= runnable_avg;
2165 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2169 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2170 int force_update) {}
2171 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2172 struct cfs_rq *cfs_rq) {}
2173 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2176 static inline void __update_task_entity_contrib(struct sched_entity *se)
2180 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2181 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2182 contrib /= (se->avg.runnable_avg_period + 1);
2183 se->avg.load_avg_contrib = scale_load(contrib);
2186 /* Compute the current contribution to load_avg by se, return any delta */
2187 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2189 long old_contrib = se->avg.load_avg_contrib;
2191 if (entity_is_task(se)) {
2192 __update_task_entity_contrib(se);
2194 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2195 __update_group_entity_contrib(se);
2198 return se->avg.load_avg_contrib - old_contrib;
2201 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2204 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2205 cfs_rq->blocked_load_avg -= load_contrib;
2207 cfs_rq->blocked_load_avg = 0;
2210 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2212 /* Update a sched_entity's runnable average */
2213 static inline void update_entity_load_avg(struct sched_entity *se,
2216 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2221 * For a group entity we need to use their owned cfs_rq_clock_task() in
2222 * case they are the parent of a throttled hierarchy.
2224 if (entity_is_task(se))
2225 now = cfs_rq_clock_task(cfs_rq);
2227 now = cfs_rq_clock_task(group_cfs_rq(se));
2229 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2232 contrib_delta = __update_entity_load_avg_contrib(se);
2238 cfs_rq->runnable_load_avg += contrib_delta;
2240 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2244 * Decay the load contributed by all blocked children and account this so that
2245 * their contribution may appropriately discounted when they wake up.
2247 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2249 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2252 decays = now - cfs_rq->last_decay;
2253 if (!decays && !force_update)
2256 if (atomic_long_read(&cfs_rq->removed_load)) {
2257 unsigned long removed_load;
2258 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2259 subtract_blocked_load_contrib(cfs_rq, removed_load);
2263 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2265 atomic64_add(decays, &cfs_rq->decay_counter);
2266 cfs_rq->last_decay = now;
2269 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2272 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2274 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2275 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2278 /* Add the load generated by se into cfs_rq's child load-average */
2279 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2280 struct sched_entity *se,
2284 * We track migrations using entity decay_count <= 0, on a wake-up
2285 * migration we use a negative decay count to track the remote decays
2286 * accumulated while sleeping.
2288 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2289 * are seen by enqueue_entity_load_avg() as a migration with an already
2290 * constructed load_avg_contrib.
2292 if (unlikely(se->avg.decay_count <= 0)) {
2293 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2294 if (se->avg.decay_count) {
2296 * In a wake-up migration we have to approximate the
2297 * time sleeping. This is because we can't synchronize
2298 * clock_task between the two cpus, and it is not
2299 * guaranteed to be read-safe. Instead, we can
2300 * approximate this using our carried decays, which are
2301 * explicitly atomically readable.
2303 se->avg.last_runnable_update -= (-se->avg.decay_count)
2305 update_entity_load_avg(se, 0);
2306 /* Indicate that we're now synchronized and on-rq */
2307 se->avg.decay_count = 0;
2312 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2313 * would have made count negative); we must be careful to avoid
2314 * double-accounting blocked time after synchronizing decays.
2316 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2320 /* migrated tasks did not contribute to our blocked load */
2322 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2323 update_entity_load_avg(se, 0);
2326 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2327 /* we force update consideration on load-balancer moves */
2328 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2332 * Remove se's load from this cfs_rq child load-average, if the entity is
2333 * transitioning to a blocked state we track its projected decay using
2336 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2337 struct sched_entity *se,
2340 update_entity_load_avg(se, 1);
2341 /* we force update consideration on load-balancer moves */
2342 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2344 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2346 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2347 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2348 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2352 * Update the rq's load with the elapsed running time before entering
2353 * idle. if the last scheduled task is not a CFS task, idle_enter will
2354 * be the only way to update the runnable statistic.
2356 void idle_enter_fair(struct rq *this_rq)
2358 update_rq_runnable_avg(this_rq, 1);
2362 * Update the rq's load with the elapsed idle time before a task is
2363 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2364 * be the only way to update the runnable statistic.
2366 void idle_exit_fair(struct rq *this_rq)
2368 update_rq_runnable_avg(this_rq, 0);
2372 static inline void update_entity_load_avg(struct sched_entity *se,
2373 int update_cfs_rq) {}
2374 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2375 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2376 struct sched_entity *se,
2378 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2379 struct sched_entity *se,
2381 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2382 int force_update) {}
2385 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2387 #ifdef CONFIG_SCHEDSTATS
2388 struct task_struct *tsk = NULL;
2390 if (entity_is_task(se))
2393 if (se->statistics.sleep_start) {
2394 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2399 if (unlikely(delta > se->statistics.sleep_max))
2400 se->statistics.sleep_max = delta;
2402 se->statistics.sleep_start = 0;
2403 se->statistics.sum_sleep_runtime += delta;
2406 account_scheduler_latency(tsk, delta >> 10, 1);
2407 trace_sched_stat_sleep(tsk, delta);
2410 if (se->statistics.block_start) {
2411 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2416 if (unlikely(delta > se->statistics.block_max))
2417 se->statistics.block_max = delta;
2419 se->statistics.block_start = 0;
2420 se->statistics.sum_sleep_runtime += delta;
2423 if (tsk->in_iowait) {
2424 se->statistics.iowait_sum += delta;
2425 se->statistics.iowait_count++;
2426 trace_sched_stat_iowait(tsk, delta);
2429 trace_sched_stat_blocked(tsk, delta);
2432 * Blocking time is in units of nanosecs, so shift by
2433 * 20 to get a milliseconds-range estimation of the
2434 * amount of time that the task spent sleeping:
2436 if (unlikely(prof_on == SLEEP_PROFILING)) {
2437 profile_hits(SLEEP_PROFILING,
2438 (void *)get_wchan(tsk),
2441 account_scheduler_latency(tsk, delta >> 10, 0);
2447 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2449 #ifdef CONFIG_SCHED_DEBUG
2450 s64 d = se->vruntime - cfs_rq->min_vruntime;
2455 if (d > 3*sysctl_sched_latency)
2456 schedstat_inc(cfs_rq, nr_spread_over);
2461 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2463 u64 vruntime = cfs_rq->min_vruntime;
2466 * The 'current' period is already promised to the current tasks,
2467 * however the extra weight of the new task will slow them down a
2468 * little, place the new task so that it fits in the slot that
2469 * stays open at the end.
2471 if (initial && sched_feat(START_DEBIT))
2472 vruntime += sched_vslice(cfs_rq, se);
2474 /* sleeps up to a single latency don't count. */
2476 unsigned long thresh = sysctl_sched_latency;
2479 * Halve their sleep time's effect, to allow
2480 * for a gentler effect of sleepers:
2482 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2488 /* ensure we never gain time by being placed backwards. */
2489 se->vruntime = max_vruntime(se->vruntime, vruntime);
2492 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2495 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2498 * Update the normalized vruntime before updating min_vruntime
2499 * through calling update_curr().
2501 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2502 se->vruntime += cfs_rq->min_vruntime;
2505 * Update run-time statistics of the 'current'.
2507 update_curr(cfs_rq);
2508 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2509 account_entity_enqueue(cfs_rq, se);
2510 update_cfs_shares(cfs_rq);
2512 if (flags & ENQUEUE_WAKEUP) {
2513 place_entity(cfs_rq, se, 0);
2514 enqueue_sleeper(cfs_rq, se);
2517 update_stats_enqueue(cfs_rq, se);
2518 check_spread(cfs_rq, se);
2519 if (se != cfs_rq->curr)
2520 __enqueue_entity(cfs_rq, se);
2523 if (cfs_rq->nr_running == 1) {
2524 list_add_leaf_cfs_rq(cfs_rq);
2525 check_enqueue_throttle(cfs_rq);
2529 static void __clear_buddies_last(struct sched_entity *se)
2531 for_each_sched_entity(se) {
2532 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2533 if (cfs_rq->last == se)
2534 cfs_rq->last = NULL;
2540 static void __clear_buddies_next(struct sched_entity *se)
2542 for_each_sched_entity(se) {
2543 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2544 if (cfs_rq->next == se)
2545 cfs_rq->next = NULL;
2551 static void __clear_buddies_skip(struct sched_entity *se)
2553 for_each_sched_entity(se) {
2554 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2555 if (cfs_rq->skip == se)
2556 cfs_rq->skip = NULL;
2562 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2564 if (cfs_rq->last == se)
2565 __clear_buddies_last(se);
2567 if (cfs_rq->next == se)
2568 __clear_buddies_next(se);
2570 if (cfs_rq->skip == se)
2571 __clear_buddies_skip(se);
2574 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2577 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2580 * Update run-time statistics of the 'current'.
2582 update_curr(cfs_rq);
2583 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2585 update_stats_dequeue(cfs_rq, se);
2586 if (flags & DEQUEUE_SLEEP) {
2587 #ifdef CONFIG_SCHEDSTATS
2588 if (entity_is_task(se)) {
2589 struct task_struct *tsk = task_of(se);
2591 if (tsk->state & TASK_INTERRUPTIBLE)
2592 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2593 if (tsk->state & TASK_UNINTERRUPTIBLE)
2594 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2599 clear_buddies(cfs_rq, se);
2601 if (se != cfs_rq->curr)
2602 __dequeue_entity(cfs_rq, se);
2604 account_entity_dequeue(cfs_rq, se);
2607 * Normalize the entity after updating the min_vruntime because the
2608 * update can refer to the ->curr item and we need to reflect this
2609 * movement in our normalized position.
2611 if (!(flags & DEQUEUE_SLEEP))
2612 se->vruntime -= cfs_rq->min_vruntime;
2614 /* return excess runtime on last dequeue */
2615 return_cfs_rq_runtime(cfs_rq);
2617 update_min_vruntime(cfs_rq);
2618 update_cfs_shares(cfs_rq);
2622 * Preempt the current task with a newly woken task if needed:
2625 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2627 unsigned long ideal_runtime, delta_exec;
2628 struct sched_entity *se;
2631 ideal_runtime = sched_slice(cfs_rq, curr);
2632 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2633 if (delta_exec > ideal_runtime) {
2634 resched_task(rq_of(cfs_rq)->curr);
2636 * The current task ran long enough, ensure it doesn't get
2637 * re-elected due to buddy favours.
2639 clear_buddies(cfs_rq, curr);
2644 * Ensure that a task that missed wakeup preemption by a
2645 * narrow margin doesn't have to wait for a full slice.
2646 * This also mitigates buddy induced latencies under load.
2648 if (delta_exec < sysctl_sched_min_granularity)
2651 se = __pick_first_entity(cfs_rq);
2652 delta = curr->vruntime - se->vruntime;
2657 if (delta > ideal_runtime)
2658 resched_task(rq_of(cfs_rq)->curr);
2662 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2664 /* 'current' is not kept within the tree. */
2667 * Any task has to be enqueued before it get to execute on
2668 * a CPU. So account for the time it spent waiting on the
2671 update_stats_wait_end(cfs_rq, se);
2672 __dequeue_entity(cfs_rq, se);
2675 update_stats_curr_start(cfs_rq, se);
2677 #ifdef CONFIG_SCHEDSTATS
2679 * Track our maximum slice length, if the CPU's load is at
2680 * least twice that of our own weight (i.e. dont track it
2681 * when there are only lesser-weight tasks around):
2683 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2684 se->statistics.slice_max = max(se->statistics.slice_max,
2685 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2688 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2692 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2695 * Pick the next process, keeping these things in mind, in this order:
2696 * 1) keep things fair between processes/task groups
2697 * 2) pick the "next" process, since someone really wants that to run
2698 * 3) pick the "last" process, for cache locality
2699 * 4) do not run the "skip" process, if something else is available
2701 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2703 struct sched_entity *se = __pick_first_entity(cfs_rq);
2704 struct sched_entity *left = se;
2707 * Avoid running the skip buddy, if running something else can
2708 * be done without getting too unfair.
2710 if (cfs_rq->skip == se) {
2711 struct sched_entity *second = __pick_next_entity(se);
2712 if (second && wakeup_preempt_entity(second, left) < 1)
2717 * Prefer last buddy, try to return the CPU to a preempted task.
2719 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2723 * Someone really wants this to run. If it's not unfair, run it.
2725 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2728 clear_buddies(cfs_rq, se);
2733 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2735 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2738 * If still on the runqueue then deactivate_task()
2739 * was not called and update_curr() has to be done:
2742 update_curr(cfs_rq);
2744 /* throttle cfs_rqs exceeding runtime */
2745 check_cfs_rq_runtime(cfs_rq);
2747 check_spread(cfs_rq, prev);
2749 update_stats_wait_start(cfs_rq, prev);
2750 /* Put 'current' back into the tree. */
2751 __enqueue_entity(cfs_rq, prev);
2752 /* in !on_rq case, update occurred at dequeue */
2753 update_entity_load_avg(prev, 1);
2755 cfs_rq->curr = NULL;
2759 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2762 * Update run-time statistics of the 'current'.
2764 update_curr(cfs_rq);
2767 * Ensure that runnable average is periodically updated.
2769 update_entity_load_avg(curr, 1);
2770 update_cfs_rq_blocked_load(cfs_rq, 1);
2771 update_cfs_shares(cfs_rq);
2773 #ifdef CONFIG_SCHED_HRTICK
2775 * queued ticks are scheduled to match the slice, so don't bother
2776 * validating it and just reschedule.
2779 resched_task(rq_of(cfs_rq)->curr);
2783 * don't let the period tick interfere with the hrtick preemption
2785 if (!sched_feat(DOUBLE_TICK) &&
2786 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2790 if (cfs_rq->nr_running > 1)
2791 check_preempt_tick(cfs_rq, curr);
2795 /**************************************************
2796 * CFS bandwidth control machinery
2799 #ifdef CONFIG_CFS_BANDWIDTH
2801 #ifdef HAVE_JUMP_LABEL
2802 static struct static_key __cfs_bandwidth_used;
2804 static inline bool cfs_bandwidth_used(void)
2806 return static_key_false(&__cfs_bandwidth_used);
2809 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2811 /* only need to count groups transitioning between enabled/!enabled */
2812 if (enabled && !was_enabled)
2813 static_key_slow_inc(&__cfs_bandwidth_used);
2814 else if (!enabled && was_enabled)
2815 static_key_slow_dec(&__cfs_bandwidth_used);
2817 #else /* HAVE_JUMP_LABEL */
2818 static bool cfs_bandwidth_used(void)
2823 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2824 #endif /* HAVE_JUMP_LABEL */
2827 * default period for cfs group bandwidth.
2828 * default: 0.1s, units: nanoseconds
2830 static inline u64 default_cfs_period(void)
2832 return 100000000ULL;
2835 static inline u64 sched_cfs_bandwidth_slice(void)
2837 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2841 * Replenish runtime according to assigned quota and update expiration time.
2842 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2843 * additional synchronization around rq->lock.
2845 * requires cfs_b->lock
2847 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2851 if (cfs_b->quota == RUNTIME_INF)
2854 now = sched_clock_cpu(smp_processor_id());
2855 cfs_b->runtime = cfs_b->quota;
2856 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2859 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2861 return &tg->cfs_bandwidth;
2864 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2865 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2867 if (unlikely(cfs_rq->throttle_count))
2868 return cfs_rq->throttled_clock_task;
2870 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2873 /* returns 0 on failure to allocate runtime */
2874 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2876 struct task_group *tg = cfs_rq->tg;
2877 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2878 u64 amount = 0, min_amount, expires;
2880 /* note: this is a positive sum as runtime_remaining <= 0 */
2881 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2883 raw_spin_lock(&cfs_b->lock);
2884 if (cfs_b->quota == RUNTIME_INF)
2885 amount = min_amount;
2888 * If the bandwidth pool has become inactive, then at least one
2889 * period must have elapsed since the last consumption.
2890 * Refresh the global state and ensure bandwidth timer becomes
2893 if (!cfs_b->timer_active) {
2894 __refill_cfs_bandwidth_runtime(cfs_b);
2895 __start_cfs_bandwidth(cfs_b);
2898 if (cfs_b->runtime > 0) {
2899 amount = min(cfs_b->runtime, min_amount);
2900 cfs_b->runtime -= amount;
2904 expires = cfs_b->runtime_expires;
2905 raw_spin_unlock(&cfs_b->lock);
2907 cfs_rq->runtime_remaining += amount;
2909 * we may have advanced our local expiration to account for allowed
2910 * spread between our sched_clock and the one on which runtime was
2913 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2914 cfs_rq->runtime_expires = expires;
2916 return cfs_rq->runtime_remaining > 0;
2920 * Note: This depends on the synchronization provided by sched_clock and the
2921 * fact that rq->clock snapshots this value.
2923 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2925 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2927 /* if the deadline is ahead of our clock, nothing to do */
2928 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2931 if (cfs_rq->runtime_remaining < 0)
2935 * If the local deadline has passed we have to consider the
2936 * possibility that our sched_clock is 'fast' and the global deadline
2937 * has not truly expired.
2939 * Fortunately we can check determine whether this the case by checking
2940 * whether the global deadline has advanced.
2943 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2944 /* extend local deadline, drift is bounded above by 2 ticks */
2945 cfs_rq->runtime_expires += TICK_NSEC;
2947 /* global deadline is ahead, expiration has passed */
2948 cfs_rq->runtime_remaining = 0;
2952 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2953 unsigned long delta_exec)
2955 /* dock delta_exec before expiring quota (as it could span periods) */
2956 cfs_rq->runtime_remaining -= delta_exec;
2957 expire_cfs_rq_runtime(cfs_rq);
2959 if (likely(cfs_rq->runtime_remaining > 0))
2963 * if we're unable to extend our runtime we resched so that the active
2964 * hierarchy can be throttled
2966 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2967 resched_task(rq_of(cfs_rq)->curr);
2970 static __always_inline
2971 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2973 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2976 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2979 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2981 return cfs_bandwidth_used() && cfs_rq->throttled;
2984 /* check whether cfs_rq, or any parent, is throttled */
2985 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2987 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2991 * Ensure that neither of the group entities corresponding to src_cpu or
2992 * dest_cpu are members of a throttled hierarchy when performing group
2993 * load-balance operations.
2995 static inline int throttled_lb_pair(struct task_group *tg,
2996 int src_cpu, int dest_cpu)
2998 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3000 src_cfs_rq = tg->cfs_rq[src_cpu];
3001 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3003 return throttled_hierarchy(src_cfs_rq) ||
3004 throttled_hierarchy(dest_cfs_rq);
3007 /* updated child weight may affect parent so we have to do this bottom up */
3008 static int tg_unthrottle_up(struct task_group *tg, void *data)
3010 struct rq *rq = data;
3011 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3013 cfs_rq->throttle_count--;
3015 if (!cfs_rq->throttle_count) {
3016 /* adjust cfs_rq_clock_task() */
3017 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3018 cfs_rq->throttled_clock_task;
3025 static int tg_throttle_down(struct task_group *tg, void *data)
3027 struct rq *rq = data;
3028 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3030 /* group is entering throttled state, stop time */
3031 if (!cfs_rq->throttle_count)
3032 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3033 cfs_rq->throttle_count++;
3038 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3040 struct rq *rq = rq_of(cfs_rq);
3041 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3042 struct sched_entity *se;
3043 long task_delta, dequeue = 1;
3045 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3047 /* freeze hierarchy runnable averages while throttled */
3049 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3052 task_delta = cfs_rq->h_nr_running;
3053 for_each_sched_entity(se) {
3054 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3055 /* throttled entity or throttle-on-deactivate */
3060 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3061 qcfs_rq->h_nr_running -= task_delta;
3063 if (qcfs_rq->load.weight)
3068 rq->nr_running -= task_delta;
3070 cfs_rq->throttled = 1;
3071 cfs_rq->throttled_clock = rq_clock(rq);
3072 raw_spin_lock(&cfs_b->lock);
3073 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3074 raw_spin_unlock(&cfs_b->lock);
3077 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3079 struct rq *rq = rq_of(cfs_rq);
3080 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3081 struct sched_entity *se;
3085 se = cfs_rq->tg->se[cpu_of(rq)];
3087 cfs_rq->throttled = 0;
3089 update_rq_clock(rq);
3091 raw_spin_lock(&cfs_b->lock);
3092 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3093 list_del_rcu(&cfs_rq->throttled_list);
3094 raw_spin_unlock(&cfs_b->lock);
3096 /* update hierarchical throttle state */
3097 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3099 if (!cfs_rq->load.weight)
3102 task_delta = cfs_rq->h_nr_running;
3103 for_each_sched_entity(se) {
3107 cfs_rq = cfs_rq_of(se);
3109 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3110 cfs_rq->h_nr_running += task_delta;
3112 if (cfs_rq_throttled(cfs_rq))
3117 rq->nr_running += task_delta;
3119 /* determine whether we need to wake up potentially idle cpu */
3120 if (rq->curr == rq->idle && rq->cfs.nr_running)
3121 resched_task(rq->curr);
3124 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3125 u64 remaining, u64 expires)
3127 struct cfs_rq *cfs_rq;
3128 u64 runtime = remaining;
3131 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3133 struct rq *rq = rq_of(cfs_rq);
3135 raw_spin_lock(&rq->lock);
3136 if (!cfs_rq_throttled(cfs_rq))
3139 runtime = -cfs_rq->runtime_remaining + 1;
3140 if (runtime > remaining)
3141 runtime = remaining;
3142 remaining -= runtime;
3144 cfs_rq->runtime_remaining += runtime;
3145 cfs_rq->runtime_expires = expires;
3147 /* we check whether we're throttled above */
3148 if (cfs_rq->runtime_remaining > 0)
3149 unthrottle_cfs_rq(cfs_rq);
3152 raw_spin_unlock(&rq->lock);
3163 * Responsible for refilling a task_group's bandwidth and unthrottling its
3164 * cfs_rqs as appropriate. If there has been no activity within the last
3165 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3166 * used to track this state.
3168 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3170 u64 runtime, runtime_expires;
3171 int idle = 1, throttled;
3173 raw_spin_lock(&cfs_b->lock);
3174 /* no need to continue the timer with no bandwidth constraint */
3175 if (cfs_b->quota == RUNTIME_INF)
3178 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3179 /* idle depends on !throttled (for the case of a large deficit) */
3180 idle = cfs_b->idle && !throttled;
3181 cfs_b->nr_periods += overrun;
3183 /* if we're going inactive then everything else can be deferred */
3187 __refill_cfs_bandwidth_runtime(cfs_b);
3190 /* mark as potentially idle for the upcoming period */
3195 /* account preceding periods in which throttling occurred */
3196 cfs_b->nr_throttled += overrun;
3199 * There are throttled entities so we must first use the new bandwidth
3200 * to unthrottle them before making it generally available. This
3201 * ensures that all existing debts will be paid before a new cfs_rq is
3204 runtime = cfs_b->runtime;
3205 runtime_expires = cfs_b->runtime_expires;
3209 * This check is repeated as we are holding onto the new bandwidth
3210 * while we unthrottle. This can potentially race with an unthrottled
3211 * group trying to acquire new bandwidth from the global pool.
3213 while (throttled && runtime > 0) {
3214 raw_spin_unlock(&cfs_b->lock);
3215 /* we can't nest cfs_b->lock while distributing bandwidth */
3216 runtime = distribute_cfs_runtime(cfs_b, runtime,
3218 raw_spin_lock(&cfs_b->lock);
3220 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3223 /* return (any) remaining runtime */
3224 cfs_b->runtime = runtime;
3226 * While we are ensured activity in the period following an
3227 * unthrottle, this also covers the case in which the new bandwidth is
3228 * insufficient to cover the existing bandwidth deficit. (Forcing the
3229 * timer to remain active while there are any throttled entities.)
3234 cfs_b->timer_active = 0;
3235 raw_spin_unlock(&cfs_b->lock);
3240 /* a cfs_rq won't donate quota below this amount */
3241 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3242 /* minimum remaining period time to redistribute slack quota */
3243 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3244 /* how long we wait to gather additional slack before distributing */
3245 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3247 /* are we near the end of the current quota period? */
3248 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3250 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3253 /* if the call-back is running a quota refresh is already occurring */
3254 if (hrtimer_callback_running(refresh_timer))
3257 /* is a quota refresh about to occur? */
3258 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3259 if (remaining < min_expire)
3265 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3267 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3269 /* if there's a quota refresh soon don't bother with slack */
3270 if (runtime_refresh_within(cfs_b, min_left))
3273 start_bandwidth_timer(&cfs_b->slack_timer,
3274 ns_to_ktime(cfs_bandwidth_slack_period));
3277 /* we know any runtime found here is valid as update_curr() precedes return */
3278 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3280 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3281 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3283 if (slack_runtime <= 0)
3286 raw_spin_lock(&cfs_b->lock);
3287 if (cfs_b->quota != RUNTIME_INF &&
3288 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3289 cfs_b->runtime += slack_runtime;
3291 /* we are under rq->lock, defer unthrottling using a timer */
3292 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3293 !list_empty(&cfs_b->throttled_cfs_rq))
3294 start_cfs_slack_bandwidth(cfs_b);
3296 raw_spin_unlock(&cfs_b->lock);
3298 /* even if it's not valid for return we don't want to try again */
3299 cfs_rq->runtime_remaining -= slack_runtime;
3302 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3304 if (!cfs_bandwidth_used())
3307 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3310 __return_cfs_rq_runtime(cfs_rq);
3314 * This is done with a timer (instead of inline with bandwidth return) since
3315 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3317 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3319 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3322 /* confirm we're still not at a refresh boundary */
3323 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3326 raw_spin_lock(&cfs_b->lock);
3327 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3328 runtime = cfs_b->runtime;
3331 expires = cfs_b->runtime_expires;
3332 raw_spin_unlock(&cfs_b->lock);
3337 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3339 raw_spin_lock(&cfs_b->lock);
3340 if (expires == cfs_b->runtime_expires)
3341 cfs_b->runtime = runtime;
3342 raw_spin_unlock(&cfs_b->lock);
3346 * When a group wakes up we want to make sure that its quota is not already
3347 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3348 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3350 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3352 if (!cfs_bandwidth_used())
3355 /* an active group must be handled by the update_curr()->put() path */
3356 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3359 /* ensure the group is not already throttled */
3360 if (cfs_rq_throttled(cfs_rq))
3363 /* update runtime allocation */
3364 account_cfs_rq_runtime(cfs_rq, 0);
3365 if (cfs_rq->runtime_remaining <= 0)
3366 throttle_cfs_rq(cfs_rq);
3369 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3370 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3372 if (!cfs_bandwidth_used())
3375 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3379 * it's possible for a throttled entity to be forced into a running
3380 * state (e.g. set_curr_task), in this case we're finished.
3382 if (cfs_rq_throttled(cfs_rq))
3385 throttle_cfs_rq(cfs_rq);
3388 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3390 struct cfs_bandwidth *cfs_b =
3391 container_of(timer, struct cfs_bandwidth, slack_timer);
3392 do_sched_cfs_slack_timer(cfs_b);
3394 return HRTIMER_NORESTART;
3397 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3399 struct cfs_bandwidth *cfs_b =
3400 container_of(timer, struct cfs_bandwidth, period_timer);
3406 now = hrtimer_cb_get_time(timer);
3407 overrun = hrtimer_forward(timer, now, cfs_b->period);
3412 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3415 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3418 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3420 raw_spin_lock_init(&cfs_b->lock);
3422 cfs_b->quota = RUNTIME_INF;
3423 cfs_b->period = ns_to_ktime(default_cfs_period());
3425 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3426 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3427 cfs_b->period_timer.function = sched_cfs_period_timer;
3428 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3429 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3432 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3434 cfs_rq->runtime_enabled = 0;
3435 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3438 /* requires cfs_b->lock, may release to reprogram timer */
3439 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3442 * The timer may be active because we're trying to set a new bandwidth
3443 * period or because we're racing with the tear-down path
3444 * (timer_active==0 becomes visible before the hrtimer call-back
3445 * terminates). In either case we ensure that it's re-programmed
3447 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3448 raw_spin_unlock(&cfs_b->lock);
3449 /* ensure cfs_b->lock is available while we wait */
3450 hrtimer_cancel(&cfs_b->period_timer);
3452 raw_spin_lock(&cfs_b->lock);
3453 /* if someone else restarted the timer then we're done */
3454 if (cfs_b->timer_active)
3458 cfs_b->timer_active = 1;
3459 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3462 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3464 hrtimer_cancel(&cfs_b->period_timer);
3465 hrtimer_cancel(&cfs_b->slack_timer);
3468 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3470 struct cfs_rq *cfs_rq;
3472 for_each_leaf_cfs_rq(rq, cfs_rq) {
3473 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3475 if (!cfs_rq->runtime_enabled)
3479 * clock_task is not advancing so we just need to make sure
3480 * there's some valid quota amount
3482 cfs_rq->runtime_remaining = cfs_b->quota;
3483 if (cfs_rq_throttled(cfs_rq))
3484 unthrottle_cfs_rq(cfs_rq);
3488 #else /* CONFIG_CFS_BANDWIDTH */
3489 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3491 return rq_clock_task(rq_of(cfs_rq));
3494 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3495 unsigned long delta_exec) {}
3496 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3497 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3498 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3500 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3505 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3510 static inline int throttled_lb_pair(struct task_group *tg,
3511 int src_cpu, int dest_cpu)
3516 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3518 #ifdef CONFIG_FAIR_GROUP_SCHED
3519 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3522 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3526 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3527 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3529 #endif /* CONFIG_CFS_BANDWIDTH */
3531 /**************************************************
3532 * CFS operations on tasks:
3535 #ifdef CONFIG_SCHED_HRTICK
3536 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3538 struct sched_entity *se = &p->se;
3539 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3541 WARN_ON(task_rq(p) != rq);
3543 if (cfs_rq->nr_running > 1) {
3544 u64 slice = sched_slice(cfs_rq, se);
3545 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3546 s64 delta = slice - ran;
3555 * Don't schedule slices shorter than 10000ns, that just
3556 * doesn't make sense. Rely on vruntime for fairness.
3559 delta = max_t(s64, 10000LL, delta);
3561 hrtick_start(rq, delta);
3566 * called from enqueue/dequeue and updates the hrtick when the
3567 * current task is from our class and nr_running is low enough
3570 static void hrtick_update(struct rq *rq)
3572 struct task_struct *curr = rq->curr;
3574 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3577 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3578 hrtick_start_fair(rq, curr);
3580 #else /* !CONFIG_SCHED_HRTICK */
3582 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3586 static inline void hrtick_update(struct rq *rq)
3592 * The enqueue_task method is called before nr_running is
3593 * increased. Here we update the fair scheduling stats and
3594 * then put the task into the rbtree:
3597 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3599 struct cfs_rq *cfs_rq;
3600 struct sched_entity *se = &p->se;
3602 for_each_sched_entity(se) {
3605 cfs_rq = cfs_rq_of(se);
3606 enqueue_entity(cfs_rq, se, flags);
3609 * end evaluation on encountering a throttled cfs_rq
3611 * note: in the case of encountering a throttled cfs_rq we will
3612 * post the final h_nr_running increment below.
3614 if (cfs_rq_throttled(cfs_rq))
3616 cfs_rq->h_nr_running++;
3618 flags = ENQUEUE_WAKEUP;
3621 for_each_sched_entity(se) {
3622 cfs_rq = cfs_rq_of(se);
3623 cfs_rq->h_nr_running++;
3625 if (cfs_rq_throttled(cfs_rq))
3628 update_cfs_shares(cfs_rq);
3629 update_entity_load_avg(se, 1);
3633 update_rq_runnable_avg(rq, rq->nr_running);
3639 static void set_next_buddy(struct sched_entity *se);
3642 * The dequeue_task method is called before nr_running is
3643 * decreased. We remove the task from the rbtree and
3644 * update the fair scheduling stats:
3646 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3648 struct cfs_rq *cfs_rq;
3649 struct sched_entity *se = &p->se;
3650 int task_sleep = flags & DEQUEUE_SLEEP;
3652 for_each_sched_entity(se) {
3653 cfs_rq = cfs_rq_of(se);
3654 dequeue_entity(cfs_rq, se, flags);
3657 * end evaluation on encountering a throttled cfs_rq
3659 * note: in the case of encountering a throttled cfs_rq we will
3660 * post the final h_nr_running decrement below.
3662 if (cfs_rq_throttled(cfs_rq))
3664 cfs_rq->h_nr_running--;
3666 /* Don't dequeue parent if it has other entities besides us */
3667 if (cfs_rq->load.weight) {
3669 * Bias pick_next to pick a task from this cfs_rq, as
3670 * p is sleeping when it is within its sched_slice.
3672 if (task_sleep && parent_entity(se))
3673 set_next_buddy(parent_entity(se));
3675 /* avoid re-evaluating load for this entity */
3676 se = parent_entity(se);
3679 flags |= DEQUEUE_SLEEP;
3682 for_each_sched_entity(se) {
3683 cfs_rq = cfs_rq_of(se);
3684 cfs_rq->h_nr_running--;
3686 if (cfs_rq_throttled(cfs_rq))
3689 update_cfs_shares(cfs_rq);
3690 update_entity_load_avg(se, 1);
3695 update_rq_runnable_avg(rq, 1);
3701 /* Used instead of source_load when we know the type == 0 */
3702 static unsigned long weighted_cpuload(const int cpu)
3704 return cpu_rq(cpu)->cfs.runnable_load_avg;
3708 * Return a low guess at the load of a migration-source cpu weighted
3709 * according to the scheduling class and "nice" value.
3711 * We want to under-estimate the load of migration sources, to
3712 * balance conservatively.
3714 static unsigned long source_load(int cpu, int type)
3716 struct rq *rq = cpu_rq(cpu);
3717 unsigned long total = weighted_cpuload(cpu);
3719 if (type == 0 || !sched_feat(LB_BIAS))
3722 return min(rq->cpu_load[type-1], total);
3726 * Return a high guess at the load of a migration-target cpu weighted
3727 * according to the scheduling class and "nice" value.
3729 static unsigned long target_load(int cpu, int type)
3731 struct rq *rq = cpu_rq(cpu);
3732 unsigned long total = weighted_cpuload(cpu);
3734 if (type == 0 || !sched_feat(LB_BIAS))
3737 return max(rq->cpu_load[type-1], total);
3740 static unsigned long power_of(int cpu)
3742 return cpu_rq(cpu)->cpu_power;
3745 static unsigned long cpu_avg_load_per_task(int cpu)
3747 struct rq *rq = cpu_rq(cpu);
3748 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3749 unsigned long load_avg = rq->cfs.runnable_load_avg;
3752 return load_avg / nr_running;
3757 static void record_wakee(struct task_struct *p)
3760 * Rough decay (wiping) for cost saving, don't worry
3761 * about the boundary, really active task won't care
3764 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3765 current->wakee_flips = 0;
3766 current->wakee_flip_decay_ts = jiffies;
3769 if (current->last_wakee != p) {
3770 current->last_wakee = p;
3771 current->wakee_flips++;
3775 static void task_waking_fair(struct task_struct *p)
3777 struct sched_entity *se = &p->se;
3778 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3781 #ifndef CONFIG_64BIT
3782 u64 min_vruntime_copy;
3785 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3787 min_vruntime = cfs_rq->min_vruntime;
3788 } while (min_vruntime != min_vruntime_copy);
3790 min_vruntime = cfs_rq->min_vruntime;
3793 se->vruntime -= min_vruntime;
3797 #ifdef CONFIG_FAIR_GROUP_SCHED
3799 * effective_load() calculates the load change as seen from the root_task_group
3801 * Adding load to a group doesn't make a group heavier, but can cause movement
3802 * of group shares between cpus. Assuming the shares were perfectly aligned one
3803 * can calculate the shift in shares.
3805 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3806 * on this @cpu and results in a total addition (subtraction) of @wg to the
3807 * total group weight.
3809 * Given a runqueue weight distribution (rw_i) we can compute a shares
3810 * distribution (s_i) using:
3812 * s_i = rw_i / \Sum rw_j (1)
3814 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3815 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3816 * shares distribution (s_i):
3818 * rw_i = { 2, 4, 1, 0 }
3819 * s_i = { 2/7, 4/7, 1/7, 0 }
3821 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3822 * task used to run on and the CPU the waker is running on), we need to
3823 * compute the effect of waking a task on either CPU and, in case of a sync
3824 * wakeup, compute the effect of the current task going to sleep.
3826 * So for a change of @wl to the local @cpu with an overall group weight change
3827 * of @wl we can compute the new shares distribution (s'_i) using:
3829 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3831 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3832 * differences in waking a task to CPU 0. The additional task changes the
3833 * weight and shares distributions like:
3835 * rw'_i = { 3, 4, 1, 0 }
3836 * s'_i = { 3/8, 4/8, 1/8, 0 }
3838 * We can then compute the difference in effective weight by using:
3840 * dw_i = S * (s'_i - s_i) (3)
3842 * Where 'S' is the group weight as seen by its parent.
3844 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3845 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3846 * 4/7) times the weight of the group.
3848 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3850 struct sched_entity *se = tg->se[cpu];
3852 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3855 for_each_sched_entity(se) {
3861 * W = @wg + \Sum rw_j
3863 W = wg + calc_tg_weight(tg, se->my_q);
3868 w = se->my_q->load.weight + wl;
3871 * wl = S * s'_i; see (2)
3874 wl = (w * tg->shares) / W;
3879 * Per the above, wl is the new se->load.weight value; since
3880 * those are clipped to [MIN_SHARES, ...) do so now. See
3881 * calc_cfs_shares().
3883 if (wl < MIN_SHARES)
3887 * wl = dw_i = S * (s'_i - s_i); see (3)
3889 wl -= se->load.weight;
3892 * Recursively apply this logic to all parent groups to compute
3893 * the final effective load change on the root group. Since
3894 * only the @tg group gets extra weight, all parent groups can
3895 * only redistribute existing shares. @wl is the shift in shares
3896 * resulting from this level per the above.
3905 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3912 static int wake_wide(struct task_struct *p)
3914 int factor = this_cpu_read(sd_llc_size);
3917 * Yeah, it's the switching-frequency, could means many wakee or
3918 * rapidly switch, use factor here will just help to automatically
3919 * adjust the loose-degree, so bigger node will lead to more pull.
3921 if (p->wakee_flips > factor) {
3923 * wakee is somewhat hot, it needs certain amount of cpu
3924 * resource, so if waker is far more hot, prefer to leave
3927 if (current->wakee_flips > (factor * p->wakee_flips))
3934 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3936 s64 this_load, load;
3937 int idx, this_cpu, prev_cpu;
3938 unsigned long tl_per_task;
3939 struct task_group *tg;
3940 unsigned long weight;
3944 * If we wake multiple tasks be careful to not bounce
3945 * ourselves around too much.
3951 this_cpu = smp_processor_id();
3952 prev_cpu = task_cpu(p);
3953 load = source_load(prev_cpu, idx);
3954 this_load = target_load(this_cpu, idx);
3957 * If sync wakeup then subtract the (maximum possible)
3958 * effect of the currently running task from the load
3959 * of the current CPU:
3962 tg = task_group(current);
3963 weight = current->se.load.weight;
3965 this_load += effective_load(tg, this_cpu, -weight, -weight);
3966 load += effective_load(tg, prev_cpu, 0, -weight);
3970 weight = p->se.load.weight;
3973 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3974 * due to the sync cause above having dropped this_load to 0, we'll
3975 * always have an imbalance, but there's really nothing you can do
3976 * about that, so that's good too.
3978 * Otherwise check if either cpus are near enough in load to allow this
3979 * task to be woken on this_cpu.
3981 if (this_load > 0) {
3982 s64 this_eff_load, prev_eff_load;
3984 this_eff_load = 100;
3985 this_eff_load *= power_of(prev_cpu);
3986 this_eff_load *= this_load +
3987 effective_load(tg, this_cpu, weight, weight);
3989 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3990 prev_eff_load *= power_of(this_cpu);
3991 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3993 balanced = this_eff_load <= prev_eff_load;
3998 * If the currently running task will sleep within
3999 * a reasonable amount of time then attract this newly
4002 if (sync && balanced)
4005 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4006 tl_per_task = cpu_avg_load_per_task(this_cpu);
4009 (this_load <= load &&
4010 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4012 * This domain has SD_WAKE_AFFINE and
4013 * p is cache cold in this domain, and
4014 * there is no bad imbalance.
4016 schedstat_inc(sd, ttwu_move_affine);
4017 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4025 * find_idlest_group finds and returns the least busy CPU group within the
4028 static struct sched_group *
4029 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4030 int this_cpu, int load_idx)
4032 struct sched_group *idlest = NULL, *group = sd->groups;
4033 unsigned long min_load = ULONG_MAX, this_load = 0;
4034 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4037 unsigned long load, avg_load;
4041 /* Skip over this group if it has no CPUs allowed */
4042 if (!cpumask_intersects(sched_group_cpus(group),
4043 tsk_cpus_allowed(p)))
4046 local_group = cpumask_test_cpu(this_cpu,
4047 sched_group_cpus(group));
4049 /* Tally up the load of all CPUs in the group */
4052 for_each_cpu(i, sched_group_cpus(group)) {
4053 /* Bias balancing toward cpus of our domain */
4055 load = source_load(i, load_idx);
4057 load = target_load(i, load_idx);
4062 /* Adjust by relative CPU power of the group */
4063 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4066 this_load = avg_load;
4067 } else if (avg_load < min_load) {
4068 min_load = avg_load;
4071 } while (group = group->next, group != sd->groups);
4073 if (!idlest || 100*this_load < imbalance*min_load)
4079 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4082 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4084 unsigned long load, min_load = ULONG_MAX;
4088 /* Traverse only the allowed CPUs */
4089 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4090 load = weighted_cpuload(i);
4092 if (load < min_load || (load == min_load && i == this_cpu)) {
4102 * Try and locate an idle CPU in the sched_domain.
4104 static int select_idle_sibling(struct task_struct *p, int target)
4106 struct sched_domain *sd;
4107 struct sched_group *sg;
4108 int i = task_cpu(p);
4110 if (idle_cpu(target))
4114 * If the prevous cpu is cache affine and idle, don't be stupid.
4116 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4120 * Otherwise, iterate the domains and find an elegible idle cpu.
4122 sd = rcu_dereference(per_cpu(sd_llc, target));
4123 for_each_lower_domain(sd) {
4126 if (!cpumask_intersects(sched_group_cpus(sg),
4127 tsk_cpus_allowed(p)))
4130 for_each_cpu(i, sched_group_cpus(sg)) {
4131 if (i == target || !idle_cpu(i))
4135 target = cpumask_first_and(sched_group_cpus(sg),
4136 tsk_cpus_allowed(p));
4140 } while (sg != sd->groups);
4147 * sched_balance_self: balance the current task (running on cpu) in domains
4148 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4151 * Balance, ie. select the least loaded group.
4153 * Returns the target CPU number, or the same CPU if no balancing is needed.
4155 * preempt must be disabled.
4158 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4160 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4161 int cpu = smp_processor_id();
4163 int want_affine = 0;
4164 int sync = wake_flags & WF_SYNC;
4166 if (p->nr_cpus_allowed == 1)
4169 if (sd_flag & SD_BALANCE_WAKE) {
4170 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4176 for_each_domain(cpu, tmp) {
4177 if (!(tmp->flags & SD_LOAD_BALANCE))
4181 * If both cpu and prev_cpu are part of this domain,
4182 * cpu is a valid SD_WAKE_AFFINE target.
4184 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4185 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4190 if (tmp->flags & sd_flag)
4195 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4198 new_cpu = select_idle_sibling(p, prev_cpu);
4203 int load_idx = sd->forkexec_idx;
4204 struct sched_group *group;
4207 if (!(sd->flags & sd_flag)) {
4212 if (sd_flag & SD_BALANCE_WAKE)
4213 load_idx = sd->wake_idx;
4215 group = find_idlest_group(sd, p, cpu, load_idx);
4221 new_cpu = find_idlest_cpu(group, p, cpu);
4222 if (new_cpu == -1 || new_cpu == cpu) {
4223 /* Now try balancing at a lower domain level of cpu */
4228 /* Now try balancing at a lower domain level of new_cpu */
4230 weight = sd->span_weight;
4232 for_each_domain(cpu, tmp) {
4233 if (weight <= tmp->span_weight)
4235 if (tmp->flags & sd_flag)
4238 /* while loop will break here if sd == NULL */
4247 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4248 * cfs_rq_of(p) references at time of call are still valid and identify the
4249 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4250 * other assumptions, including the state of rq->lock, should be made.
4253 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4255 struct sched_entity *se = &p->se;
4256 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4259 * Load tracking: accumulate removed load so that it can be processed
4260 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4261 * to blocked load iff they have a positive decay-count. It can never
4262 * be negative here since on-rq tasks have decay-count == 0.
4264 if (se->avg.decay_count) {
4265 se->avg.decay_count = -__synchronize_entity_decay(se);
4266 atomic_long_add(se->avg.load_avg_contrib,
4267 &cfs_rq->removed_load);
4270 #endif /* CONFIG_SMP */
4272 static unsigned long
4273 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4275 unsigned long gran = sysctl_sched_wakeup_granularity;
4278 * Since its curr running now, convert the gran from real-time
4279 * to virtual-time in his units.
4281 * By using 'se' instead of 'curr' we penalize light tasks, so
4282 * they get preempted easier. That is, if 'se' < 'curr' then
4283 * the resulting gran will be larger, therefore penalizing the
4284 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4285 * be smaller, again penalizing the lighter task.
4287 * This is especially important for buddies when the leftmost
4288 * task is higher priority than the buddy.
4290 return calc_delta_fair(gran, se);
4294 * Should 'se' preempt 'curr'.
4308 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4310 s64 gran, vdiff = curr->vruntime - se->vruntime;
4315 gran = wakeup_gran(curr, se);
4322 static void set_last_buddy(struct sched_entity *se)
4324 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4327 for_each_sched_entity(se)
4328 cfs_rq_of(se)->last = se;
4331 static void set_next_buddy(struct sched_entity *se)
4333 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4336 for_each_sched_entity(se)
4337 cfs_rq_of(se)->next = se;
4340 static void set_skip_buddy(struct sched_entity *se)
4342 for_each_sched_entity(se)
4343 cfs_rq_of(se)->skip = se;
4347 * Preempt the current task with a newly woken task if needed:
4349 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4351 struct task_struct *curr = rq->curr;
4352 struct sched_entity *se = &curr->se, *pse = &p->se;
4353 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4354 int scale = cfs_rq->nr_running >= sched_nr_latency;
4355 int next_buddy_marked = 0;
4357 if (unlikely(se == pse))
4361 * This is possible from callers such as move_task(), in which we
4362 * unconditionally check_prempt_curr() after an enqueue (which may have
4363 * lead to a throttle). This both saves work and prevents false
4364 * next-buddy nomination below.
4366 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4369 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4370 set_next_buddy(pse);
4371 next_buddy_marked = 1;
4375 * We can come here with TIF_NEED_RESCHED already set from new task
4378 * Note: this also catches the edge-case of curr being in a throttled
4379 * group (e.g. via set_curr_task), since update_curr() (in the
4380 * enqueue of curr) will have resulted in resched being set. This
4381 * prevents us from potentially nominating it as a false LAST_BUDDY
4384 if (test_tsk_need_resched(curr))
4387 /* Idle tasks are by definition preempted by non-idle tasks. */
4388 if (unlikely(curr->policy == SCHED_IDLE) &&
4389 likely(p->policy != SCHED_IDLE))
4393 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4394 * is driven by the tick):
4396 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4399 find_matching_se(&se, &pse);
4400 update_curr(cfs_rq_of(se));
4402 if (wakeup_preempt_entity(se, pse) == 1) {
4404 * Bias pick_next to pick the sched entity that is
4405 * triggering this preemption.
4407 if (!next_buddy_marked)
4408 set_next_buddy(pse);
4417 * Only set the backward buddy when the current task is still
4418 * on the rq. This can happen when a wakeup gets interleaved
4419 * with schedule on the ->pre_schedule() or idle_balance()
4420 * point, either of which can * drop the rq lock.
4422 * Also, during early boot the idle thread is in the fair class,
4423 * for obvious reasons its a bad idea to schedule back to it.
4425 if (unlikely(!se->on_rq || curr == rq->idle))
4428 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4432 static struct task_struct *pick_next_task_fair(struct rq *rq)
4434 struct task_struct *p;
4435 struct cfs_rq *cfs_rq = &rq->cfs;
4436 struct sched_entity *se;
4438 if (!cfs_rq->nr_running)
4442 se = pick_next_entity(cfs_rq);
4443 set_next_entity(cfs_rq, se);
4444 cfs_rq = group_cfs_rq(se);
4448 if (hrtick_enabled(rq))
4449 hrtick_start_fair(rq, p);
4455 * Account for a descheduled task:
4457 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4459 struct sched_entity *se = &prev->se;
4460 struct cfs_rq *cfs_rq;
4462 for_each_sched_entity(se) {
4463 cfs_rq = cfs_rq_of(se);
4464 put_prev_entity(cfs_rq, se);
4469 * sched_yield() is very simple
4471 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4473 static void yield_task_fair(struct rq *rq)
4475 struct task_struct *curr = rq->curr;
4476 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4477 struct sched_entity *se = &curr->se;
4480 * Are we the only task in the tree?
4482 if (unlikely(rq->nr_running == 1))
4485 clear_buddies(cfs_rq, se);
4487 if (curr->policy != SCHED_BATCH) {
4488 update_rq_clock(rq);
4490 * Update run-time statistics of the 'current'.
4492 update_curr(cfs_rq);
4494 * Tell update_rq_clock() that we've just updated,
4495 * so we don't do microscopic update in schedule()
4496 * and double the fastpath cost.
4498 rq->skip_clock_update = 1;
4504 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4506 struct sched_entity *se = &p->se;
4508 /* throttled hierarchies are not runnable */
4509 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4512 /* Tell the scheduler that we'd really like pse to run next. */
4515 yield_task_fair(rq);
4521 /**************************************************
4522 * Fair scheduling class load-balancing methods.
4526 * The purpose of load-balancing is to achieve the same basic fairness the
4527 * per-cpu scheduler provides, namely provide a proportional amount of compute
4528 * time to each task. This is expressed in the following equation:
4530 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4532 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4533 * W_i,0 is defined as:
4535 * W_i,0 = \Sum_j w_i,j (2)
4537 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4538 * is derived from the nice value as per prio_to_weight[].
4540 * The weight average is an exponential decay average of the instantaneous
4543 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4545 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4546 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4547 * can also include other factors [XXX].
4549 * To achieve this balance we define a measure of imbalance which follows
4550 * directly from (1):
4552 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4554 * We them move tasks around to minimize the imbalance. In the continuous
4555 * function space it is obvious this converges, in the discrete case we get
4556 * a few fun cases generally called infeasible weight scenarios.
4559 * - infeasible weights;
4560 * - local vs global optima in the discrete case. ]
4565 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4566 * for all i,j solution, we create a tree of cpus that follows the hardware
4567 * topology where each level pairs two lower groups (or better). This results
4568 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4569 * tree to only the first of the previous level and we decrease the frequency
4570 * of load-balance at each level inv. proportional to the number of cpus in
4576 * \Sum { --- * --- * 2^i } = O(n) (5)
4578 * `- size of each group
4579 * | | `- number of cpus doing load-balance
4581 * `- sum over all levels
4583 * Coupled with a limit on how many tasks we can migrate every balance pass,
4584 * this makes (5) the runtime complexity of the balancer.
4586 * An important property here is that each CPU is still (indirectly) connected
4587 * to every other cpu in at most O(log n) steps:
4589 * The adjacency matrix of the resulting graph is given by:
4592 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4595 * And you'll find that:
4597 * A^(log_2 n)_i,j != 0 for all i,j (7)
4599 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4600 * The task movement gives a factor of O(m), giving a convergence complexity
4603 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4608 * In order to avoid CPUs going idle while there's still work to do, new idle
4609 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4610 * tree itself instead of relying on other CPUs to bring it work.
4612 * This adds some complexity to both (5) and (8) but it reduces the total idle
4620 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4623 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4628 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4630 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4632 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4635 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4636 * rewrite all of this once again.]
4639 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4641 enum fbq_type { regular, remote, all };
4643 #define LBF_ALL_PINNED 0x01
4644 #define LBF_NEED_BREAK 0x02
4645 #define LBF_DST_PINNED 0x04
4646 #define LBF_SOME_PINNED 0x08
4649 struct sched_domain *sd;
4657 struct cpumask *dst_grpmask;
4659 enum cpu_idle_type idle;
4661 /* The set of CPUs under consideration for load-balancing */
4662 struct cpumask *cpus;
4667 unsigned int loop_break;
4668 unsigned int loop_max;
4670 enum fbq_type fbq_type;
4674 * move_task - move a task from one runqueue to another runqueue.
4675 * Both runqueues must be locked.
4677 static void move_task(struct task_struct *p, struct lb_env *env)
4679 deactivate_task(env->src_rq, p, 0);
4680 set_task_cpu(p, env->dst_cpu);
4681 activate_task(env->dst_rq, p, 0);
4682 check_preempt_curr(env->dst_rq, p, 0);
4683 #ifdef CONFIG_NUMA_BALANCING
4684 if (p->numa_preferred_nid != -1) {
4685 int src_nid = cpu_to_node(env->src_cpu);
4686 int dst_nid = cpu_to_node(env->dst_cpu);
4689 * If the load balancer has moved the task then limit
4690 * migrations from taking place in the short term in
4691 * case this is a short-lived migration.
4693 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4694 p->numa_migrate_seq = 0;
4700 * Is this task likely cache-hot:
4703 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4707 if (p->sched_class != &fair_sched_class)
4710 if (unlikely(p->policy == SCHED_IDLE))
4714 * Buddy candidates are cache hot:
4716 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4717 (&p->se == cfs_rq_of(&p->se)->next ||
4718 &p->se == cfs_rq_of(&p->se)->last))
4721 if (sysctl_sched_migration_cost == -1)
4723 if (sysctl_sched_migration_cost == 0)
4726 delta = now - p->se.exec_start;
4728 return delta < (s64)sysctl_sched_migration_cost;
4731 #ifdef CONFIG_NUMA_BALANCING
4732 /* Returns true if the destination node has incurred more faults */
4733 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4735 int src_nid, dst_nid;
4737 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4738 !(env->sd->flags & SD_NUMA)) {
4742 src_nid = cpu_to_node(env->src_cpu);
4743 dst_nid = cpu_to_node(env->dst_cpu);
4745 if (src_nid == dst_nid)
4748 /* Always encourage migration to the preferred node. */
4749 if (dst_nid == p->numa_preferred_nid)
4752 /* If both task and group weight improve, this move is a winner. */
4753 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4754 group_weight(p, dst_nid) > group_weight(p, src_nid))
4761 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4763 int src_nid, dst_nid;
4765 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4768 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4771 src_nid = cpu_to_node(env->src_cpu);
4772 dst_nid = cpu_to_node(env->dst_cpu);
4774 if (src_nid == dst_nid)
4777 /* Migrating away from the preferred node is always bad. */
4778 if (src_nid == p->numa_preferred_nid)
4781 /* If either task or group weight get worse, don't do it. */
4782 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4783 group_weight(p, dst_nid) < group_weight(p, src_nid))
4790 static inline bool migrate_improves_locality(struct task_struct *p,
4796 static inline bool migrate_degrades_locality(struct task_struct *p,
4804 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4807 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4809 int tsk_cache_hot = 0;
4811 * We do not migrate tasks that are:
4812 * 1) throttled_lb_pair, or
4813 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4814 * 3) running (obviously), or
4815 * 4) are cache-hot on their current CPU.
4817 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4820 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4823 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4825 env->flags |= LBF_SOME_PINNED;
4828 * Remember if this task can be migrated to any other cpu in
4829 * our sched_group. We may want to revisit it if we couldn't
4830 * meet load balance goals by pulling other tasks on src_cpu.
4832 * Also avoid computing new_dst_cpu if we have already computed
4833 * one in current iteration.
4835 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4838 /* Prevent to re-select dst_cpu via env's cpus */
4839 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4840 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4841 env->flags |= LBF_DST_PINNED;
4842 env->new_dst_cpu = cpu;
4850 /* Record that we found atleast one task that could run on dst_cpu */
4851 env->flags &= ~LBF_ALL_PINNED;
4853 if (task_running(env->src_rq, p)) {
4854 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4859 * Aggressive migration if:
4860 * 1) destination numa is preferred
4861 * 2) task is cache cold, or
4862 * 3) too many balance attempts have failed.
4864 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4866 tsk_cache_hot = migrate_degrades_locality(p, env);
4868 if (migrate_improves_locality(p, env)) {
4869 #ifdef CONFIG_SCHEDSTATS
4870 if (tsk_cache_hot) {
4871 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4872 schedstat_inc(p, se.statistics.nr_forced_migrations);
4878 if (!tsk_cache_hot ||
4879 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4881 if (tsk_cache_hot) {
4882 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4883 schedstat_inc(p, se.statistics.nr_forced_migrations);
4889 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4894 * move_one_task tries to move exactly one task from busiest to this_rq, as
4895 * part of active balancing operations within "domain".
4896 * Returns 1 if successful and 0 otherwise.
4898 * Called with both runqueues locked.
4900 static int move_one_task(struct lb_env *env)
4902 struct task_struct *p, *n;
4904 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4905 if (!can_migrate_task(p, env))
4910 * Right now, this is only the second place move_task()
4911 * is called, so we can safely collect move_task()
4912 * stats here rather than inside move_task().
4914 schedstat_inc(env->sd, lb_gained[env->idle]);
4920 static const unsigned int sched_nr_migrate_break = 32;
4923 * move_tasks tries to move up to imbalance weighted load from busiest to
4924 * this_rq, as part of a balancing operation within domain "sd".
4925 * Returns 1 if successful and 0 otherwise.
4927 * Called with both runqueues locked.
4929 static int move_tasks(struct lb_env *env)
4931 struct list_head *tasks = &env->src_rq->cfs_tasks;
4932 struct task_struct *p;
4936 if (env->imbalance <= 0)
4939 while (!list_empty(tasks)) {
4940 p = list_first_entry(tasks, struct task_struct, se.group_node);
4943 /* We've more or less seen every task there is, call it quits */
4944 if (env->loop > env->loop_max)
4947 /* take a breather every nr_migrate tasks */
4948 if (env->loop > env->loop_break) {
4949 env->loop_break += sched_nr_migrate_break;
4950 env->flags |= LBF_NEED_BREAK;
4954 if (!can_migrate_task(p, env))
4957 load = task_h_load(p);
4959 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4962 if ((load / 2) > env->imbalance)
4967 env->imbalance -= load;
4969 #ifdef CONFIG_PREEMPT
4971 * NEWIDLE balancing is a source of latency, so preemptible
4972 * kernels will stop after the first task is pulled to minimize
4973 * the critical section.
4975 if (env->idle == CPU_NEWLY_IDLE)
4980 * We only want to steal up to the prescribed amount of
4983 if (env->imbalance <= 0)
4988 list_move_tail(&p->se.group_node, tasks);
4992 * Right now, this is one of only two places move_task() is called,
4993 * so we can safely collect move_task() stats here rather than
4994 * inside move_task().
4996 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5001 #ifdef CONFIG_FAIR_GROUP_SCHED
5003 * update tg->load_weight by folding this cpu's load_avg
5005 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5007 struct sched_entity *se = tg->se[cpu];
5008 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5010 /* throttled entities do not contribute to load */
5011 if (throttled_hierarchy(cfs_rq))
5014 update_cfs_rq_blocked_load(cfs_rq, 1);
5017 update_entity_load_avg(se, 1);
5019 * We pivot on our runnable average having decayed to zero for
5020 * list removal. This generally implies that all our children
5021 * have also been removed (modulo rounding error or bandwidth
5022 * control); however, such cases are rare and we can fix these
5025 * TODO: fix up out-of-order children on enqueue.
5027 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5028 list_del_leaf_cfs_rq(cfs_rq);
5030 struct rq *rq = rq_of(cfs_rq);
5031 update_rq_runnable_avg(rq, rq->nr_running);
5035 static void update_blocked_averages(int cpu)
5037 struct rq *rq = cpu_rq(cpu);
5038 struct cfs_rq *cfs_rq;
5039 unsigned long flags;
5041 raw_spin_lock_irqsave(&rq->lock, flags);
5042 update_rq_clock(rq);
5044 * Iterates the task_group tree in a bottom up fashion, see
5045 * list_add_leaf_cfs_rq() for details.
5047 for_each_leaf_cfs_rq(rq, cfs_rq) {
5049 * Note: We may want to consider periodically releasing
5050 * rq->lock about these updates so that creating many task
5051 * groups does not result in continually extending hold time.
5053 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5056 raw_spin_unlock_irqrestore(&rq->lock, flags);
5060 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5061 * This needs to be done in a top-down fashion because the load of a child
5062 * group is a fraction of its parents load.
5064 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5066 struct rq *rq = rq_of(cfs_rq);
5067 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5068 unsigned long now = jiffies;
5071 if (cfs_rq->last_h_load_update == now)
5074 cfs_rq->h_load_next = NULL;
5075 for_each_sched_entity(se) {
5076 cfs_rq = cfs_rq_of(se);
5077 cfs_rq->h_load_next = se;
5078 if (cfs_rq->last_h_load_update == now)
5083 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5084 cfs_rq->last_h_load_update = now;
5087 while ((se = cfs_rq->h_load_next) != NULL) {
5088 load = cfs_rq->h_load;
5089 load = div64_ul(load * se->avg.load_avg_contrib,
5090 cfs_rq->runnable_load_avg + 1);
5091 cfs_rq = group_cfs_rq(se);
5092 cfs_rq->h_load = load;
5093 cfs_rq->last_h_load_update = now;
5097 static unsigned long task_h_load(struct task_struct *p)
5099 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5101 update_cfs_rq_h_load(cfs_rq);
5102 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5103 cfs_rq->runnable_load_avg + 1);
5106 static inline void update_blocked_averages(int cpu)
5110 static unsigned long task_h_load(struct task_struct *p)
5112 return p->se.avg.load_avg_contrib;
5116 /********** Helpers for find_busiest_group ************************/
5118 * sg_lb_stats - stats of a sched_group required for load_balancing
5120 struct sg_lb_stats {
5121 unsigned long avg_load; /*Avg load across the CPUs of the group */
5122 unsigned long group_load; /* Total load over the CPUs of the group */
5123 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5124 unsigned long load_per_task;
5125 unsigned long group_power;
5126 unsigned int sum_nr_running; /* Nr tasks running in the group */
5127 unsigned int group_capacity;
5128 unsigned int idle_cpus;
5129 unsigned int group_weight;
5130 int group_imb; /* Is there an imbalance in the group ? */
5131 int group_has_capacity; /* Is there extra capacity in the group? */
5132 #ifdef CONFIG_NUMA_BALANCING
5133 unsigned int nr_numa_running;
5134 unsigned int nr_preferred_running;
5139 * sd_lb_stats - Structure to store the statistics of a sched_domain
5140 * during load balancing.
5142 struct sd_lb_stats {
5143 struct sched_group *busiest; /* Busiest group in this sd */
5144 struct sched_group *local; /* Local group in this sd */
5145 unsigned long total_load; /* Total load of all groups in sd */
5146 unsigned long total_pwr; /* Total power of all groups in sd */
5147 unsigned long avg_load; /* Average load across all groups in sd */
5149 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5150 struct sg_lb_stats local_stat; /* Statistics of the local group */
5153 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5156 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5157 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5158 * We must however clear busiest_stat::avg_load because
5159 * update_sd_pick_busiest() reads this before assignment.
5161 *sds = (struct sd_lb_stats){
5173 * get_sd_load_idx - Obtain the load index for a given sched domain.
5174 * @sd: The sched_domain whose load_idx is to be obtained.
5175 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5177 * Return: The load index.
5179 static inline int get_sd_load_idx(struct sched_domain *sd,
5180 enum cpu_idle_type idle)
5186 load_idx = sd->busy_idx;
5189 case CPU_NEWLY_IDLE:
5190 load_idx = sd->newidle_idx;
5193 load_idx = sd->idle_idx;
5200 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5202 return SCHED_POWER_SCALE;
5205 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5207 return default_scale_freq_power(sd, cpu);
5210 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5212 unsigned long weight = sd->span_weight;
5213 unsigned long smt_gain = sd->smt_gain;
5220 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5222 return default_scale_smt_power(sd, cpu);
5225 static unsigned long scale_rt_power(int cpu)
5227 struct rq *rq = cpu_rq(cpu);
5228 u64 total, available, age_stamp, avg;
5231 * Since we're reading these variables without serialization make sure
5232 * we read them once before doing sanity checks on them.
5234 age_stamp = ACCESS_ONCE(rq->age_stamp);
5235 avg = ACCESS_ONCE(rq->rt_avg);
5237 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5239 if (unlikely(total < avg)) {
5240 /* Ensures that power won't end up being negative */
5243 available = total - avg;
5246 if (unlikely((s64)total < SCHED_POWER_SCALE))
5247 total = SCHED_POWER_SCALE;
5249 total >>= SCHED_POWER_SHIFT;
5251 return div_u64(available, total);
5254 static void update_cpu_power(struct sched_domain *sd, int cpu)
5256 unsigned long weight = sd->span_weight;
5257 unsigned long power = SCHED_POWER_SCALE;
5258 struct sched_group *sdg = sd->groups;
5260 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5261 if (sched_feat(ARCH_POWER))
5262 power *= arch_scale_smt_power(sd, cpu);
5264 power *= default_scale_smt_power(sd, cpu);
5266 power >>= SCHED_POWER_SHIFT;
5269 sdg->sgp->power_orig = power;
5271 if (sched_feat(ARCH_POWER))
5272 power *= arch_scale_freq_power(sd, cpu);
5274 power *= default_scale_freq_power(sd, cpu);
5276 power >>= SCHED_POWER_SHIFT;
5278 power *= scale_rt_power(cpu);
5279 power >>= SCHED_POWER_SHIFT;
5284 cpu_rq(cpu)->cpu_power = power;
5285 sdg->sgp->power = power;
5288 void update_group_power(struct sched_domain *sd, int cpu)
5290 struct sched_domain *child = sd->child;
5291 struct sched_group *group, *sdg = sd->groups;
5292 unsigned long power, power_orig;
5293 unsigned long interval;
5295 interval = msecs_to_jiffies(sd->balance_interval);
5296 interval = clamp(interval, 1UL, max_load_balance_interval);
5297 sdg->sgp->next_update = jiffies + interval;
5300 update_cpu_power(sd, cpu);
5304 power_orig = power = 0;
5306 if (child->flags & SD_OVERLAP) {
5308 * SD_OVERLAP domains cannot assume that child groups
5309 * span the current group.
5312 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5313 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5315 power_orig += sg->sgp->power_orig;
5316 power += sg->sgp->power;
5320 * !SD_OVERLAP domains can assume that child groups
5321 * span the current group.
5324 group = child->groups;
5326 power_orig += group->sgp->power_orig;
5327 power += group->sgp->power;
5328 group = group->next;
5329 } while (group != child->groups);
5332 sdg->sgp->power_orig = power_orig;
5333 sdg->sgp->power = power;
5337 * Try and fix up capacity for tiny siblings, this is needed when
5338 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5339 * which on its own isn't powerful enough.
5341 * See update_sd_pick_busiest() and check_asym_packing().
5344 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5347 * Only siblings can have significantly less than SCHED_POWER_SCALE
5349 if (!(sd->flags & SD_SHARE_CPUPOWER))
5353 * If ~90% of the cpu_power is still there, we're good.
5355 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5362 * Group imbalance indicates (and tries to solve) the problem where balancing
5363 * groups is inadequate due to tsk_cpus_allowed() constraints.
5365 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5366 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5369 * { 0 1 2 3 } { 4 5 6 7 }
5372 * If we were to balance group-wise we'd place two tasks in the first group and
5373 * two tasks in the second group. Clearly this is undesired as it will overload
5374 * cpu 3 and leave one of the cpus in the second group unused.
5376 * The current solution to this issue is detecting the skew in the first group
5377 * by noticing the lower domain failed to reach balance and had difficulty
5378 * moving tasks due to affinity constraints.
5380 * When this is so detected; this group becomes a candidate for busiest; see
5381 * update_sd_pick_busiest(). And calculcate_imbalance() and
5382 * find_busiest_group() avoid some of the usual balance conditions to allow it
5383 * to create an effective group imbalance.
5385 * This is a somewhat tricky proposition since the next run might not find the
5386 * group imbalance and decide the groups need to be balanced again. A most
5387 * subtle and fragile situation.
5390 static inline int sg_imbalanced(struct sched_group *group)
5392 return group->sgp->imbalance;
5396 * Compute the group capacity.
5398 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5399 * first dividing out the smt factor and computing the actual number of cores
5400 * and limit power unit capacity with that.
5402 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5404 unsigned int capacity, smt, cpus;
5405 unsigned int power, power_orig;
5407 power = group->sgp->power;
5408 power_orig = group->sgp->power_orig;
5409 cpus = group->group_weight;
5411 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5412 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5413 capacity = cpus / smt; /* cores */
5415 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5417 capacity = fix_small_capacity(env->sd, group);
5423 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5424 * @env: The load balancing environment.
5425 * @group: sched_group whose statistics are to be updated.
5426 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5427 * @local_group: Does group contain this_cpu.
5428 * @sgs: variable to hold the statistics for this group.
5430 static inline void update_sg_lb_stats(struct lb_env *env,
5431 struct sched_group *group, int load_idx,
5432 int local_group, struct sg_lb_stats *sgs)
5434 unsigned long nr_running;
5438 memset(sgs, 0, sizeof(*sgs));
5440 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5441 struct rq *rq = cpu_rq(i);
5443 nr_running = rq->nr_running;
5445 /* Bias balancing toward cpus of our domain */
5447 load = target_load(i, load_idx);
5449 load = source_load(i, load_idx);
5451 sgs->group_load += load;
5452 sgs->sum_nr_running += nr_running;
5453 #ifdef CONFIG_NUMA_BALANCING
5454 sgs->nr_numa_running += rq->nr_numa_running;
5455 sgs->nr_preferred_running += rq->nr_preferred_running;
5457 sgs->sum_weighted_load += weighted_cpuload(i);
5462 /* Adjust by relative CPU power of the group */
5463 sgs->group_power = group->sgp->power;
5464 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5466 if (sgs->sum_nr_running)
5467 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5469 sgs->group_weight = group->group_weight;
5471 sgs->group_imb = sg_imbalanced(group);
5472 sgs->group_capacity = sg_capacity(env, group);
5474 if (sgs->group_capacity > sgs->sum_nr_running)
5475 sgs->group_has_capacity = 1;
5479 * update_sd_pick_busiest - return 1 on busiest group
5480 * @env: The load balancing environment.
5481 * @sds: sched_domain statistics
5482 * @sg: sched_group candidate to be checked for being the busiest
5483 * @sgs: sched_group statistics
5485 * Determine if @sg is a busier group than the previously selected
5488 * Return: %true if @sg is a busier group than the previously selected
5489 * busiest group. %false otherwise.
5491 static bool update_sd_pick_busiest(struct lb_env *env,
5492 struct sd_lb_stats *sds,
5493 struct sched_group *sg,
5494 struct sg_lb_stats *sgs)
5496 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5499 if (sgs->sum_nr_running > sgs->group_capacity)
5506 * ASYM_PACKING needs to move all the work to the lowest
5507 * numbered CPUs in the group, therefore mark all groups
5508 * higher than ourself as busy.
5510 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5511 env->dst_cpu < group_first_cpu(sg)) {
5515 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5522 #ifdef CONFIG_NUMA_BALANCING
5523 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5525 if (sgs->sum_nr_running > sgs->nr_numa_running)
5527 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5532 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5534 if (rq->nr_running > rq->nr_numa_running)
5536 if (rq->nr_running > rq->nr_preferred_running)
5541 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5546 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5550 #endif /* CONFIG_NUMA_BALANCING */
5553 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5554 * @env: The load balancing environment.
5555 * @balance: Should we balance.
5556 * @sds: variable to hold the statistics for this sched_domain.
5558 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5560 struct sched_domain *child = env->sd->child;
5561 struct sched_group *sg = env->sd->groups;
5562 struct sg_lb_stats tmp_sgs;
5563 int load_idx, prefer_sibling = 0;
5565 if (child && child->flags & SD_PREFER_SIBLING)
5568 load_idx = get_sd_load_idx(env->sd, env->idle);
5571 struct sg_lb_stats *sgs = &tmp_sgs;
5574 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5577 sgs = &sds->local_stat;
5579 if (env->idle != CPU_NEWLY_IDLE ||
5580 time_after_eq(jiffies, sg->sgp->next_update))
5581 update_group_power(env->sd, env->dst_cpu);
5584 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5590 * In case the child domain prefers tasks go to siblings
5591 * first, lower the sg capacity to one so that we'll try
5592 * and move all the excess tasks away. We lower the capacity
5593 * of a group only if the local group has the capacity to fit
5594 * these excess tasks, i.e. nr_running < group_capacity. The
5595 * extra check prevents the case where you always pull from the
5596 * heaviest group when it is already under-utilized (possible
5597 * with a large weight task outweighs the tasks on the system).
5599 if (prefer_sibling && sds->local &&
5600 sds->local_stat.group_has_capacity)
5601 sgs->group_capacity = min(sgs->group_capacity, 1U);
5603 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5605 sds->busiest_stat = *sgs;
5609 /* Now, start updating sd_lb_stats */
5610 sds->total_load += sgs->group_load;
5611 sds->total_pwr += sgs->group_power;
5614 } while (sg != env->sd->groups);
5616 if (env->sd->flags & SD_NUMA)
5617 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5621 * check_asym_packing - Check to see if the group is packed into the
5624 * This is primarily intended to used at the sibling level. Some
5625 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5626 * case of POWER7, it can move to lower SMT modes only when higher
5627 * threads are idle. When in lower SMT modes, the threads will
5628 * perform better since they share less core resources. Hence when we
5629 * have idle threads, we want them to be the higher ones.
5631 * This packing function is run on idle threads. It checks to see if
5632 * the busiest CPU in this domain (core in the P7 case) has a higher
5633 * CPU number than the packing function is being run on. Here we are
5634 * assuming lower CPU number will be equivalent to lower a SMT thread
5637 * Return: 1 when packing is required and a task should be moved to
5638 * this CPU. The amount of the imbalance is returned in *imbalance.
5640 * @env: The load balancing environment.
5641 * @sds: Statistics of the sched_domain which is to be packed
5643 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5647 if (!(env->sd->flags & SD_ASYM_PACKING))
5653 busiest_cpu = group_first_cpu(sds->busiest);
5654 if (env->dst_cpu > busiest_cpu)
5657 env->imbalance = DIV_ROUND_CLOSEST(
5658 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5665 * fix_small_imbalance - Calculate the minor imbalance that exists
5666 * amongst the groups of a sched_domain, during
5668 * @env: The load balancing environment.
5669 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5672 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5674 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5675 unsigned int imbn = 2;
5676 unsigned long scaled_busy_load_per_task;
5677 struct sg_lb_stats *local, *busiest;
5679 local = &sds->local_stat;
5680 busiest = &sds->busiest_stat;
5682 if (!local->sum_nr_running)
5683 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5684 else if (busiest->load_per_task > local->load_per_task)
5687 scaled_busy_load_per_task =
5688 (busiest->load_per_task * SCHED_POWER_SCALE) /
5689 busiest->group_power;
5691 if (busiest->avg_load + scaled_busy_load_per_task >=
5692 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5693 env->imbalance = busiest->load_per_task;
5698 * OK, we don't have enough imbalance to justify moving tasks,
5699 * however we may be able to increase total CPU power used by
5703 pwr_now += busiest->group_power *
5704 min(busiest->load_per_task, busiest->avg_load);
5705 pwr_now += local->group_power *
5706 min(local->load_per_task, local->avg_load);
5707 pwr_now /= SCHED_POWER_SCALE;
5709 /* Amount of load we'd subtract */
5710 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5711 busiest->group_power;
5712 if (busiest->avg_load > tmp) {
5713 pwr_move += busiest->group_power *
5714 min(busiest->load_per_task,
5715 busiest->avg_load - tmp);
5718 /* Amount of load we'd add */
5719 if (busiest->avg_load * busiest->group_power <
5720 busiest->load_per_task * SCHED_POWER_SCALE) {
5721 tmp = (busiest->avg_load * busiest->group_power) /
5724 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5727 pwr_move += local->group_power *
5728 min(local->load_per_task, local->avg_load + tmp);
5729 pwr_move /= SCHED_POWER_SCALE;
5731 /* Move if we gain throughput */
5732 if (pwr_move > pwr_now)
5733 env->imbalance = busiest->load_per_task;
5737 * calculate_imbalance - Calculate the amount of imbalance present within the
5738 * groups of a given sched_domain during load balance.
5739 * @env: load balance environment
5740 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5742 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5744 unsigned long max_pull, load_above_capacity = ~0UL;
5745 struct sg_lb_stats *local, *busiest;
5747 local = &sds->local_stat;
5748 busiest = &sds->busiest_stat;
5750 if (busiest->group_imb) {
5752 * In the group_imb case we cannot rely on group-wide averages
5753 * to ensure cpu-load equilibrium, look at wider averages. XXX
5755 busiest->load_per_task =
5756 min(busiest->load_per_task, sds->avg_load);
5760 * In the presence of smp nice balancing, certain scenarios can have
5761 * max load less than avg load(as we skip the groups at or below
5762 * its cpu_power, while calculating max_load..)
5764 if (busiest->avg_load <= sds->avg_load ||
5765 local->avg_load >= sds->avg_load) {
5767 return fix_small_imbalance(env, sds);
5770 if (!busiest->group_imb) {
5772 * Don't want to pull so many tasks that a group would go idle.
5773 * Except of course for the group_imb case, since then we might
5774 * have to drop below capacity to reach cpu-load equilibrium.
5776 load_above_capacity =
5777 (busiest->sum_nr_running - busiest->group_capacity);
5779 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5780 load_above_capacity /= busiest->group_power;
5784 * We're trying to get all the cpus to the average_load, so we don't
5785 * want to push ourselves above the average load, nor do we wish to
5786 * reduce the max loaded cpu below the average load. At the same time,
5787 * we also don't want to reduce the group load below the group capacity
5788 * (so that we can implement power-savings policies etc). Thus we look
5789 * for the minimum possible imbalance.
5791 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5793 /* How much load to actually move to equalise the imbalance */
5794 env->imbalance = min(
5795 max_pull * busiest->group_power,
5796 (sds->avg_load - local->avg_load) * local->group_power
5797 ) / SCHED_POWER_SCALE;
5800 * if *imbalance is less than the average load per runnable task
5801 * there is no guarantee that any tasks will be moved so we'll have
5802 * a think about bumping its value to force at least one task to be
5805 if (env->imbalance < busiest->load_per_task)
5806 return fix_small_imbalance(env, sds);
5809 /******* find_busiest_group() helpers end here *********************/
5812 * find_busiest_group - Returns the busiest group within the sched_domain
5813 * if there is an imbalance. If there isn't an imbalance, and
5814 * the user has opted for power-savings, it returns a group whose
5815 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5816 * such a group exists.
5818 * Also calculates the amount of weighted load which should be moved
5819 * to restore balance.
5821 * @env: The load balancing environment.
5823 * Return: - The busiest group if imbalance exists.
5824 * - If no imbalance and user has opted for power-savings balance,
5825 * return the least loaded group whose CPUs can be
5826 * put to idle by rebalancing its tasks onto our group.
5828 static struct sched_group *find_busiest_group(struct lb_env *env)
5830 struct sg_lb_stats *local, *busiest;
5831 struct sd_lb_stats sds;
5833 init_sd_lb_stats(&sds);
5836 * Compute the various statistics relavent for load balancing at
5839 update_sd_lb_stats(env, &sds);
5840 local = &sds.local_stat;
5841 busiest = &sds.busiest_stat;
5843 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5844 check_asym_packing(env, &sds))
5847 /* There is no busy sibling group to pull tasks from */
5848 if (!sds.busiest || busiest->sum_nr_running == 0)
5851 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5854 * If the busiest group is imbalanced the below checks don't
5855 * work because they assume all things are equal, which typically
5856 * isn't true due to cpus_allowed constraints and the like.
5858 if (busiest->group_imb)
5861 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5862 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5863 !busiest->group_has_capacity)
5867 * If the local group is more busy than the selected busiest group
5868 * don't try and pull any tasks.
5870 if (local->avg_load >= busiest->avg_load)
5874 * Don't pull any tasks if this group is already above the domain
5877 if (local->avg_load >= sds.avg_load)
5880 if (env->idle == CPU_IDLE) {
5882 * This cpu is idle. If the busiest group load doesn't
5883 * have more tasks than the number of available cpu's and
5884 * there is no imbalance between this and busiest group
5885 * wrt to idle cpu's, it is balanced.
5887 if ((local->idle_cpus < busiest->idle_cpus) &&
5888 busiest->sum_nr_running <= busiest->group_weight)
5892 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5893 * imbalance_pct to be conservative.
5895 if (100 * busiest->avg_load <=
5896 env->sd->imbalance_pct * local->avg_load)
5901 /* Looks like there is an imbalance. Compute it */
5902 calculate_imbalance(env, &sds);
5911 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5913 static struct rq *find_busiest_queue(struct lb_env *env,
5914 struct sched_group *group)
5916 struct rq *busiest = NULL, *rq;
5917 unsigned long busiest_load = 0, busiest_power = 1;
5920 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5921 unsigned long power, capacity, wl;
5925 rt = fbq_classify_rq(rq);
5928 * We classify groups/runqueues into three groups:
5929 * - regular: there are !numa tasks
5930 * - remote: there are numa tasks that run on the 'wrong' node
5931 * - all: there is no distinction
5933 * In order to avoid migrating ideally placed numa tasks,
5934 * ignore those when there's better options.
5936 * If we ignore the actual busiest queue to migrate another
5937 * task, the next balance pass can still reduce the busiest
5938 * queue by moving tasks around inside the node.
5940 * If we cannot move enough load due to this classification
5941 * the next pass will adjust the group classification and
5942 * allow migration of more tasks.
5944 * Both cases only affect the total convergence complexity.
5946 if (rt > env->fbq_type)
5949 power = power_of(i);
5950 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5952 capacity = fix_small_capacity(env->sd, group);
5954 wl = weighted_cpuload(i);
5957 * When comparing with imbalance, use weighted_cpuload()
5958 * which is not scaled with the cpu power.
5960 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5964 * For the load comparisons with the other cpu's, consider
5965 * the weighted_cpuload() scaled with the cpu power, so that
5966 * the load can be moved away from the cpu that is potentially
5967 * running at a lower capacity.
5969 * Thus we're looking for max(wl_i / power_i), crosswise
5970 * multiplication to rid ourselves of the division works out
5971 * to: wl_i * power_j > wl_j * power_i; where j is our
5974 if (wl * busiest_power > busiest_load * power) {
5976 busiest_power = power;
5985 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5986 * so long as it is large enough.
5988 #define MAX_PINNED_INTERVAL 512
5990 /* Working cpumask for load_balance and load_balance_newidle. */
5991 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5993 static int need_active_balance(struct lb_env *env)
5995 struct sched_domain *sd = env->sd;
5997 if (env->idle == CPU_NEWLY_IDLE) {
6000 * ASYM_PACKING needs to force migrate tasks from busy but
6001 * higher numbered CPUs in order to pack all tasks in the
6002 * lowest numbered CPUs.
6004 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6008 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6011 static int active_load_balance_cpu_stop(void *data);
6013 static int should_we_balance(struct lb_env *env)
6015 struct sched_group *sg = env->sd->groups;
6016 struct cpumask *sg_cpus, *sg_mask;
6017 int cpu, balance_cpu = -1;
6020 * In the newly idle case, we will allow all the cpu's
6021 * to do the newly idle load balance.
6023 if (env->idle == CPU_NEWLY_IDLE)
6026 sg_cpus = sched_group_cpus(sg);
6027 sg_mask = sched_group_mask(sg);
6028 /* Try to find first idle cpu */
6029 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6030 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6037 if (balance_cpu == -1)
6038 balance_cpu = group_balance_cpu(sg);
6041 * First idle cpu or the first cpu(busiest) in this sched group
6042 * is eligible for doing load balancing at this and above domains.
6044 return balance_cpu == env->dst_cpu;
6048 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6049 * tasks if there is an imbalance.
6051 static int load_balance(int this_cpu, struct rq *this_rq,
6052 struct sched_domain *sd, enum cpu_idle_type idle,
6053 int *continue_balancing)
6055 int ld_moved, cur_ld_moved, active_balance = 0;
6056 struct sched_domain *sd_parent = sd->parent;
6057 struct sched_group *group;
6059 unsigned long flags;
6060 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6062 struct lb_env env = {
6064 .dst_cpu = this_cpu,
6066 .dst_grpmask = sched_group_cpus(sd->groups),
6068 .loop_break = sched_nr_migrate_break,
6074 * For NEWLY_IDLE load_balancing, we don't need to consider
6075 * other cpus in our group
6077 if (idle == CPU_NEWLY_IDLE)
6078 env.dst_grpmask = NULL;
6080 cpumask_copy(cpus, cpu_active_mask);
6082 schedstat_inc(sd, lb_count[idle]);
6085 if (!should_we_balance(&env)) {
6086 *continue_balancing = 0;
6090 group = find_busiest_group(&env);
6092 schedstat_inc(sd, lb_nobusyg[idle]);
6096 busiest = find_busiest_queue(&env, group);
6098 schedstat_inc(sd, lb_nobusyq[idle]);
6102 BUG_ON(busiest == env.dst_rq);
6104 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6107 if (busiest->nr_running > 1) {
6109 * Attempt to move tasks. If find_busiest_group has found
6110 * an imbalance but busiest->nr_running <= 1, the group is
6111 * still unbalanced. ld_moved simply stays zero, so it is
6112 * correctly treated as an imbalance.
6114 env.flags |= LBF_ALL_PINNED;
6115 env.src_cpu = busiest->cpu;
6116 env.src_rq = busiest;
6117 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6120 local_irq_save(flags);
6121 double_rq_lock(env.dst_rq, busiest);
6124 * cur_ld_moved - load moved in current iteration
6125 * ld_moved - cumulative load moved across iterations
6127 cur_ld_moved = move_tasks(&env);
6128 ld_moved += cur_ld_moved;
6129 double_rq_unlock(env.dst_rq, busiest);
6130 local_irq_restore(flags);
6133 * some other cpu did the load balance for us.
6135 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6136 resched_cpu(env.dst_cpu);
6138 if (env.flags & LBF_NEED_BREAK) {
6139 env.flags &= ~LBF_NEED_BREAK;
6144 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6145 * us and move them to an alternate dst_cpu in our sched_group
6146 * where they can run. The upper limit on how many times we
6147 * iterate on same src_cpu is dependent on number of cpus in our
6150 * This changes load balance semantics a bit on who can move
6151 * load to a given_cpu. In addition to the given_cpu itself
6152 * (or a ilb_cpu acting on its behalf where given_cpu is
6153 * nohz-idle), we now have balance_cpu in a position to move
6154 * load to given_cpu. In rare situations, this may cause
6155 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6156 * _independently_ and at _same_ time to move some load to
6157 * given_cpu) causing exceess load to be moved to given_cpu.
6158 * This however should not happen so much in practice and
6159 * moreover subsequent load balance cycles should correct the
6160 * excess load moved.
6162 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6164 /* Prevent to re-select dst_cpu via env's cpus */
6165 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6167 env.dst_rq = cpu_rq(env.new_dst_cpu);
6168 env.dst_cpu = env.new_dst_cpu;
6169 env.flags &= ~LBF_DST_PINNED;
6171 env.loop_break = sched_nr_migrate_break;
6174 * Go back to "more_balance" rather than "redo" since we
6175 * need to continue with same src_cpu.
6181 * We failed to reach balance because of affinity.
6184 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6186 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6187 *group_imbalance = 1;
6188 } else if (*group_imbalance)
6189 *group_imbalance = 0;
6192 /* All tasks on this runqueue were pinned by CPU affinity */
6193 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6194 cpumask_clear_cpu(cpu_of(busiest), cpus);
6195 if (!cpumask_empty(cpus)) {
6197 env.loop_break = sched_nr_migrate_break;
6205 schedstat_inc(sd, lb_failed[idle]);
6207 * Increment the failure counter only on periodic balance.
6208 * We do not want newidle balance, which can be very
6209 * frequent, pollute the failure counter causing
6210 * excessive cache_hot migrations and active balances.
6212 if (idle != CPU_NEWLY_IDLE)
6213 sd->nr_balance_failed++;
6215 if (need_active_balance(&env)) {
6216 raw_spin_lock_irqsave(&busiest->lock, flags);
6218 /* don't kick the active_load_balance_cpu_stop,
6219 * if the curr task on busiest cpu can't be
6222 if (!cpumask_test_cpu(this_cpu,
6223 tsk_cpus_allowed(busiest->curr))) {
6224 raw_spin_unlock_irqrestore(&busiest->lock,
6226 env.flags |= LBF_ALL_PINNED;
6227 goto out_one_pinned;
6231 * ->active_balance synchronizes accesses to
6232 * ->active_balance_work. Once set, it's cleared
6233 * only after active load balance is finished.
6235 if (!busiest->active_balance) {
6236 busiest->active_balance = 1;
6237 busiest->push_cpu = this_cpu;
6240 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6242 if (active_balance) {
6243 stop_one_cpu_nowait(cpu_of(busiest),
6244 active_load_balance_cpu_stop, busiest,
6245 &busiest->active_balance_work);
6249 * We've kicked active balancing, reset the failure
6252 sd->nr_balance_failed = sd->cache_nice_tries+1;
6255 sd->nr_balance_failed = 0;
6257 if (likely(!active_balance)) {
6258 /* We were unbalanced, so reset the balancing interval */
6259 sd->balance_interval = sd->min_interval;
6262 * If we've begun active balancing, start to back off. This
6263 * case may not be covered by the all_pinned logic if there
6264 * is only 1 task on the busy runqueue (because we don't call
6267 if (sd->balance_interval < sd->max_interval)
6268 sd->balance_interval *= 2;
6274 schedstat_inc(sd, lb_balanced[idle]);
6276 sd->nr_balance_failed = 0;
6279 /* tune up the balancing interval */
6280 if (((env.flags & LBF_ALL_PINNED) &&
6281 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6282 (sd->balance_interval < sd->max_interval))
6283 sd->balance_interval *= 2;
6291 * idle_balance is called by schedule() if this_cpu is about to become
6292 * idle. Attempts to pull tasks from other CPUs.
6294 void idle_balance(int this_cpu, struct rq *this_rq)
6296 struct sched_domain *sd;
6297 int pulled_task = 0;
6298 unsigned long next_balance = jiffies + HZ;
6301 this_rq->idle_stamp = rq_clock(this_rq);
6303 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6307 * Drop the rq->lock, but keep IRQ/preempt disabled.
6309 raw_spin_unlock(&this_rq->lock);
6311 update_blocked_averages(this_cpu);
6313 for_each_domain(this_cpu, sd) {
6314 unsigned long interval;
6315 int continue_balancing = 1;
6316 u64 t0, domain_cost;
6318 if (!(sd->flags & SD_LOAD_BALANCE))
6321 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6324 if (sd->flags & SD_BALANCE_NEWIDLE) {
6325 t0 = sched_clock_cpu(this_cpu);
6327 /* If we've pulled tasks over stop searching: */
6328 pulled_task = load_balance(this_cpu, this_rq,
6330 &continue_balancing);
6332 domain_cost = sched_clock_cpu(this_cpu) - t0;
6333 if (domain_cost > sd->max_newidle_lb_cost)
6334 sd->max_newidle_lb_cost = domain_cost;
6336 curr_cost += domain_cost;
6339 interval = msecs_to_jiffies(sd->balance_interval);
6340 if (time_after(next_balance, sd->last_balance + interval))
6341 next_balance = sd->last_balance + interval;
6343 this_rq->idle_stamp = 0;
6349 raw_spin_lock(&this_rq->lock);
6351 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6353 * We are going idle. next_balance may be set based on
6354 * a busy processor. So reset next_balance.
6356 this_rq->next_balance = next_balance;
6359 if (curr_cost > this_rq->max_idle_balance_cost)
6360 this_rq->max_idle_balance_cost = curr_cost;
6364 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6365 * running tasks off the busiest CPU onto idle CPUs. It requires at
6366 * least 1 task to be running on each physical CPU where possible, and
6367 * avoids physical / logical imbalances.
6369 static int active_load_balance_cpu_stop(void *data)
6371 struct rq *busiest_rq = data;
6372 int busiest_cpu = cpu_of(busiest_rq);
6373 int target_cpu = busiest_rq->push_cpu;
6374 struct rq *target_rq = cpu_rq(target_cpu);
6375 struct sched_domain *sd;
6377 raw_spin_lock_irq(&busiest_rq->lock);
6379 /* make sure the requested cpu hasn't gone down in the meantime */
6380 if (unlikely(busiest_cpu != smp_processor_id() ||
6381 !busiest_rq->active_balance))
6384 /* Is there any task to move? */
6385 if (busiest_rq->nr_running <= 1)
6389 * This condition is "impossible", if it occurs
6390 * we need to fix it. Originally reported by
6391 * Bjorn Helgaas on a 128-cpu setup.
6393 BUG_ON(busiest_rq == target_rq);
6395 /* move a task from busiest_rq to target_rq */
6396 double_lock_balance(busiest_rq, target_rq);
6398 /* Search for an sd spanning us and the target CPU. */
6400 for_each_domain(target_cpu, sd) {
6401 if ((sd->flags & SD_LOAD_BALANCE) &&
6402 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6407 struct lb_env env = {
6409 .dst_cpu = target_cpu,
6410 .dst_rq = target_rq,
6411 .src_cpu = busiest_rq->cpu,
6412 .src_rq = busiest_rq,
6416 schedstat_inc(sd, alb_count);
6418 if (move_one_task(&env))
6419 schedstat_inc(sd, alb_pushed);
6421 schedstat_inc(sd, alb_failed);
6424 double_unlock_balance(busiest_rq, target_rq);
6426 busiest_rq->active_balance = 0;
6427 raw_spin_unlock_irq(&busiest_rq->lock);
6431 #ifdef CONFIG_NO_HZ_COMMON
6433 * idle load balancing details
6434 * - When one of the busy CPUs notice that there may be an idle rebalancing
6435 * needed, they will kick the idle load balancer, which then does idle
6436 * load balancing for all the idle CPUs.
6439 cpumask_var_t idle_cpus_mask;
6441 unsigned long next_balance; /* in jiffy units */
6442 } nohz ____cacheline_aligned;
6444 static inline int find_new_ilb(int call_cpu)
6446 int ilb = cpumask_first(nohz.idle_cpus_mask);
6448 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6455 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6456 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6457 * CPU (if there is one).
6459 static void nohz_balancer_kick(int cpu)
6463 nohz.next_balance++;
6465 ilb_cpu = find_new_ilb(cpu);
6467 if (ilb_cpu >= nr_cpu_ids)
6470 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6473 * Use smp_send_reschedule() instead of resched_cpu().
6474 * This way we generate a sched IPI on the target cpu which
6475 * is idle. And the softirq performing nohz idle load balance
6476 * will be run before returning from the IPI.
6478 smp_send_reschedule(ilb_cpu);
6482 static inline void nohz_balance_exit_idle(int cpu)
6484 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6485 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6486 atomic_dec(&nohz.nr_cpus);
6487 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6491 static inline void set_cpu_sd_state_busy(void)
6493 struct sched_domain *sd;
6496 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6498 if (!sd || !sd->nohz_idle)
6502 for (; sd; sd = sd->parent)
6503 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6508 void set_cpu_sd_state_idle(void)
6510 struct sched_domain *sd;
6513 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6515 if (!sd || sd->nohz_idle)
6519 for (; sd; sd = sd->parent)
6520 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6526 * This routine will record that the cpu is going idle with tick stopped.
6527 * This info will be used in performing idle load balancing in the future.
6529 void nohz_balance_enter_idle(int cpu)
6532 * If this cpu is going down, then nothing needs to be done.
6534 if (!cpu_active(cpu))
6537 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6540 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6541 atomic_inc(&nohz.nr_cpus);
6542 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6545 static int sched_ilb_notifier(struct notifier_block *nfb,
6546 unsigned long action, void *hcpu)
6548 switch (action & ~CPU_TASKS_FROZEN) {
6550 nohz_balance_exit_idle(smp_processor_id());
6558 static DEFINE_SPINLOCK(balancing);
6561 * Scale the max load_balance interval with the number of CPUs in the system.
6562 * This trades load-balance latency on larger machines for less cross talk.
6564 void update_max_interval(void)
6566 max_load_balance_interval = HZ*num_online_cpus()/10;
6570 * It checks each scheduling domain to see if it is due to be balanced,
6571 * and initiates a balancing operation if so.
6573 * Balancing parameters are set up in init_sched_domains.
6575 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6577 int continue_balancing = 1;
6578 struct rq *rq = cpu_rq(cpu);
6579 unsigned long interval;
6580 struct sched_domain *sd;
6581 /* Earliest time when we have to do rebalance again */
6582 unsigned long next_balance = jiffies + 60*HZ;
6583 int update_next_balance = 0;
6584 int need_serialize, need_decay = 0;
6587 update_blocked_averages(cpu);
6590 for_each_domain(cpu, sd) {
6592 * Decay the newidle max times here because this is a regular
6593 * visit to all the domains. Decay ~1% per second.
6595 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6596 sd->max_newidle_lb_cost =
6597 (sd->max_newidle_lb_cost * 253) / 256;
6598 sd->next_decay_max_lb_cost = jiffies + HZ;
6601 max_cost += sd->max_newidle_lb_cost;
6603 if (!(sd->flags & SD_LOAD_BALANCE))
6607 * Stop the load balance at this level. There is another
6608 * CPU in our sched group which is doing load balancing more
6611 if (!continue_balancing) {
6617 interval = sd->balance_interval;
6618 if (idle != CPU_IDLE)
6619 interval *= sd->busy_factor;
6621 /* scale ms to jiffies */
6622 interval = msecs_to_jiffies(interval);
6623 interval = clamp(interval, 1UL, max_load_balance_interval);
6625 need_serialize = sd->flags & SD_SERIALIZE;
6627 if (need_serialize) {
6628 if (!spin_trylock(&balancing))
6632 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6633 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6635 * The LBF_DST_PINNED logic could have changed
6636 * env->dst_cpu, so we can't know our idle
6637 * state even if we migrated tasks. Update it.
6639 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6641 sd->last_balance = jiffies;
6644 spin_unlock(&balancing);
6646 if (time_after(next_balance, sd->last_balance + interval)) {
6647 next_balance = sd->last_balance + interval;
6648 update_next_balance = 1;
6653 * Ensure the rq-wide value also decays but keep it at a
6654 * reasonable floor to avoid funnies with rq->avg_idle.
6656 rq->max_idle_balance_cost =
6657 max((u64)sysctl_sched_migration_cost, max_cost);
6662 * next_balance will be updated only when there is a need.
6663 * When the cpu is attached to null domain for ex, it will not be
6666 if (likely(update_next_balance))
6667 rq->next_balance = next_balance;
6670 #ifdef CONFIG_NO_HZ_COMMON
6672 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6673 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6675 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6677 struct rq *this_rq = cpu_rq(this_cpu);
6681 if (idle != CPU_IDLE ||
6682 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6685 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6686 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6690 * If this cpu gets work to do, stop the load balancing
6691 * work being done for other cpus. Next load
6692 * balancing owner will pick it up.
6697 rq = cpu_rq(balance_cpu);
6699 raw_spin_lock_irq(&rq->lock);
6700 update_rq_clock(rq);
6701 update_idle_cpu_load(rq);
6702 raw_spin_unlock_irq(&rq->lock);
6704 rebalance_domains(balance_cpu, CPU_IDLE);
6706 if (time_after(this_rq->next_balance, rq->next_balance))
6707 this_rq->next_balance = rq->next_balance;
6709 nohz.next_balance = this_rq->next_balance;
6711 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6715 * Current heuristic for kicking the idle load balancer in the presence
6716 * of an idle cpu is the system.
6717 * - This rq has more than one task.
6718 * - At any scheduler domain level, this cpu's scheduler group has multiple
6719 * busy cpu's exceeding the group's power.
6720 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6721 * domain span are idle.
6723 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6725 unsigned long now = jiffies;
6726 struct sched_domain *sd;
6728 if (unlikely(idle_cpu(cpu)))
6732 * We may be recently in ticked or tickless idle mode. At the first
6733 * busy tick after returning from idle, we will update the busy stats.
6735 set_cpu_sd_state_busy();
6736 nohz_balance_exit_idle(cpu);
6739 * None are in tickless mode and hence no need for NOHZ idle load
6742 if (likely(!atomic_read(&nohz.nr_cpus)))
6745 if (time_before(now, nohz.next_balance))
6748 if (rq->nr_running >= 2)
6752 for_each_domain(cpu, sd) {
6753 struct sched_group *sg = sd->groups;
6754 struct sched_group_power *sgp = sg->sgp;
6755 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6757 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6758 goto need_kick_unlock;
6760 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6761 && (cpumask_first_and(nohz.idle_cpus_mask,
6762 sched_domain_span(sd)) < cpu))
6763 goto need_kick_unlock;
6765 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6777 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6781 * run_rebalance_domains is triggered when needed from the scheduler tick.
6782 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6784 static void run_rebalance_domains(struct softirq_action *h)
6786 int this_cpu = smp_processor_id();
6787 struct rq *this_rq = cpu_rq(this_cpu);
6788 enum cpu_idle_type idle = this_rq->idle_balance ?
6789 CPU_IDLE : CPU_NOT_IDLE;
6791 rebalance_domains(this_cpu, idle);
6794 * If this cpu has a pending nohz_balance_kick, then do the
6795 * balancing on behalf of the other idle cpus whose ticks are
6798 nohz_idle_balance(this_cpu, idle);
6801 static inline int on_null_domain(int cpu)
6803 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6807 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6809 void trigger_load_balance(struct rq *rq, int cpu)
6811 /* Don't need to rebalance while attached to NULL domain */
6812 if (time_after_eq(jiffies, rq->next_balance) &&
6813 likely(!on_null_domain(cpu)))
6814 raise_softirq(SCHED_SOFTIRQ);
6815 #ifdef CONFIG_NO_HZ_COMMON
6816 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6817 nohz_balancer_kick(cpu);
6821 static void rq_online_fair(struct rq *rq)
6826 static void rq_offline_fair(struct rq *rq)
6830 /* Ensure any throttled groups are reachable by pick_next_task */
6831 unthrottle_offline_cfs_rqs(rq);
6834 #endif /* CONFIG_SMP */
6837 * scheduler tick hitting a task of our scheduling class:
6839 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6841 struct cfs_rq *cfs_rq;
6842 struct sched_entity *se = &curr->se;
6844 for_each_sched_entity(se) {
6845 cfs_rq = cfs_rq_of(se);
6846 entity_tick(cfs_rq, se, queued);
6849 if (numabalancing_enabled)
6850 task_tick_numa(rq, curr);
6852 update_rq_runnable_avg(rq, 1);
6856 * called on fork with the child task as argument from the parent's context
6857 * - child not yet on the tasklist
6858 * - preemption disabled
6860 static void task_fork_fair(struct task_struct *p)
6862 struct cfs_rq *cfs_rq;
6863 struct sched_entity *se = &p->se, *curr;
6864 int this_cpu = smp_processor_id();
6865 struct rq *rq = this_rq();
6866 unsigned long flags;
6868 raw_spin_lock_irqsave(&rq->lock, flags);
6870 update_rq_clock(rq);
6872 cfs_rq = task_cfs_rq(current);
6873 curr = cfs_rq->curr;
6876 * Not only the cpu but also the task_group of the parent might have
6877 * been changed after parent->se.parent,cfs_rq were copied to
6878 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6879 * of child point to valid ones.
6882 __set_task_cpu(p, this_cpu);
6885 update_curr(cfs_rq);
6888 se->vruntime = curr->vruntime;
6889 place_entity(cfs_rq, se, 1);
6891 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6893 * Upon rescheduling, sched_class::put_prev_task() will place
6894 * 'current' within the tree based on its new key value.
6896 swap(curr->vruntime, se->vruntime);
6897 resched_task(rq->curr);
6900 se->vruntime -= cfs_rq->min_vruntime;
6902 raw_spin_unlock_irqrestore(&rq->lock, flags);
6906 * Priority of the task has changed. Check to see if we preempt
6910 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6916 * Reschedule if we are currently running on this runqueue and
6917 * our priority decreased, or if we are not currently running on
6918 * this runqueue and our priority is higher than the current's
6920 if (rq->curr == p) {
6921 if (p->prio > oldprio)
6922 resched_task(rq->curr);
6924 check_preempt_curr(rq, p, 0);
6927 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6929 struct sched_entity *se = &p->se;
6930 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6933 * Ensure the task's vruntime is normalized, so that when its
6934 * switched back to the fair class the enqueue_entity(.flags=0) will
6935 * do the right thing.
6937 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6938 * have normalized the vruntime, if it was !on_rq, then only when
6939 * the task is sleeping will it still have non-normalized vruntime.
6941 if (!se->on_rq && p->state != TASK_RUNNING) {
6943 * Fix up our vruntime so that the current sleep doesn't
6944 * cause 'unlimited' sleep bonus.
6946 place_entity(cfs_rq, se, 0);
6947 se->vruntime -= cfs_rq->min_vruntime;
6952 * Remove our load from contribution when we leave sched_fair
6953 * and ensure we don't carry in an old decay_count if we
6956 if (se->avg.decay_count) {
6957 __synchronize_entity_decay(se);
6958 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6964 * We switched to the sched_fair class.
6966 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6972 * We were most likely switched from sched_rt, so
6973 * kick off the schedule if running, otherwise just see
6974 * if we can still preempt the current task.
6977 resched_task(rq->curr);
6979 check_preempt_curr(rq, p, 0);
6982 /* Account for a task changing its policy or group.
6984 * This routine is mostly called to set cfs_rq->curr field when a task
6985 * migrates between groups/classes.
6987 static void set_curr_task_fair(struct rq *rq)
6989 struct sched_entity *se = &rq->curr->se;
6991 for_each_sched_entity(se) {
6992 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6994 set_next_entity(cfs_rq, se);
6995 /* ensure bandwidth has been allocated on our new cfs_rq */
6996 account_cfs_rq_runtime(cfs_rq, 0);
7000 void init_cfs_rq(struct cfs_rq *cfs_rq)
7002 cfs_rq->tasks_timeline = RB_ROOT;
7003 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7004 #ifndef CONFIG_64BIT
7005 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7008 atomic64_set(&cfs_rq->decay_counter, 1);
7009 atomic_long_set(&cfs_rq->removed_load, 0);
7013 #ifdef CONFIG_FAIR_GROUP_SCHED
7014 static void task_move_group_fair(struct task_struct *p, int on_rq)
7016 struct cfs_rq *cfs_rq;
7018 * If the task was not on the rq at the time of this cgroup movement
7019 * it must have been asleep, sleeping tasks keep their ->vruntime
7020 * absolute on their old rq until wakeup (needed for the fair sleeper
7021 * bonus in place_entity()).
7023 * If it was on the rq, we've just 'preempted' it, which does convert
7024 * ->vruntime to a relative base.
7026 * Make sure both cases convert their relative position when migrating
7027 * to another cgroup's rq. This does somewhat interfere with the
7028 * fair sleeper stuff for the first placement, but who cares.
7031 * When !on_rq, vruntime of the task has usually NOT been normalized.
7032 * But there are some cases where it has already been normalized:
7034 * - Moving a forked child which is waiting for being woken up by
7035 * wake_up_new_task().
7036 * - Moving a task which has been woken up by try_to_wake_up() and
7037 * waiting for actually being woken up by sched_ttwu_pending().
7039 * To prevent boost or penalty in the new cfs_rq caused by delta
7040 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7042 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7046 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7047 set_task_rq(p, task_cpu(p));
7049 cfs_rq = cfs_rq_of(&p->se);
7050 p->se.vruntime += cfs_rq->min_vruntime;
7053 * migrate_task_rq_fair() will have removed our previous
7054 * contribution, but we must synchronize for ongoing future
7057 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7058 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7063 void free_fair_sched_group(struct task_group *tg)
7067 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7069 for_each_possible_cpu(i) {
7071 kfree(tg->cfs_rq[i]);
7080 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7082 struct cfs_rq *cfs_rq;
7083 struct sched_entity *se;
7086 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7089 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7093 tg->shares = NICE_0_LOAD;
7095 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7097 for_each_possible_cpu(i) {
7098 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7099 GFP_KERNEL, cpu_to_node(i));
7103 se = kzalloc_node(sizeof(struct sched_entity),
7104 GFP_KERNEL, cpu_to_node(i));
7108 init_cfs_rq(cfs_rq);
7109 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7120 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7122 struct rq *rq = cpu_rq(cpu);
7123 unsigned long flags;
7126 * Only empty task groups can be destroyed; so we can speculatively
7127 * check on_list without danger of it being re-added.
7129 if (!tg->cfs_rq[cpu]->on_list)
7132 raw_spin_lock_irqsave(&rq->lock, flags);
7133 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7134 raw_spin_unlock_irqrestore(&rq->lock, flags);
7137 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7138 struct sched_entity *se, int cpu,
7139 struct sched_entity *parent)
7141 struct rq *rq = cpu_rq(cpu);
7145 init_cfs_rq_runtime(cfs_rq);
7147 tg->cfs_rq[cpu] = cfs_rq;
7150 /* se could be NULL for root_task_group */
7155 se->cfs_rq = &rq->cfs;
7157 se->cfs_rq = parent->my_q;
7160 update_load_set(&se->load, 0);
7161 se->parent = parent;
7164 static DEFINE_MUTEX(shares_mutex);
7166 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7169 unsigned long flags;
7172 * We can't change the weight of the root cgroup.
7177 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7179 mutex_lock(&shares_mutex);
7180 if (tg->shares == shares)
7183 tg->shares = shares;
7184 for_each_possible_cpu(i) {
7185 struct rq *rq = cpu_rq(i);
7186 struct sched_entity *se;
7189 /* Propagate contribution to hierarchy */
7190 raw_spin_lock_irqsave(&rq->lock, flags);
7192 /* Possible calls to update_curr() need rq clock */
7193 update_rq_clock(rq);
7194 for_each_sched_entity(se)
7195 update_cfs_shares(group_cfs_rq(se));
7196 raw_spin_unlock_irqrestore(&rq->lock, flags);
7200 mutex_unlock(&shares_mutex);
7203 #else /* CONFIG_FAIR_GROUP_SCHED */
7205 void free_fair_sched_group(struct task_group *tg) { }
7207 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7212 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7214 #endif /* CONFIG_FAIR_GROUP_SCHED */
7217 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7219 struct sched_entity *se = &task->se;
7220 unsigned int rr_interval = 0;
7223 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7226 if (rq->cfs.load.weight)
7227 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7233 * All the scheduling class methods:
7235 const struct sched_class fair_sched_class = {
7236 .next = &idle_sched_class,
7237 .enqueue_task = enqueue_task_fair,
7238 .dequeue_task = dequeue_task_fair,
7239 .yield_task = yield_task_fair,
7240 .yield_to_task = yield_to_task_fair,
7242 .check_preempt_curr = check_preempt_wakeup,
7244 .pick_next_task = pick_next_task_fair,
7245 .put_prev_task = put_prev_task_fair,
7248 .select_task_rq = select_task_rq_fair,
7249 .migrate_task_rq = migrate_task_rq_fair,
7251 .rq_online = rq_online_fair,
7252 .rq_offline = rq_offline_fair,
7254 .task_waking = task_waking_fair,
7257 .set_curr_task = set_curr_task_fair,
7258 .task_tick = task_tick_fair,
7259 .task_fork = task_fork_fair,
7261 .prio_changed = prio_changed_fair,
7262 .switched_from = switched_from_fair,
7263 .switched_to = switched_to_fair,
7265 .get_rr_interval = get_rr_interval_fair,
7267 #ifdef CONFIG_FAIR_GROUP_SCHED
7268 .task_move_group = task_move_group_fair,
7272 #ifdef CONFIG_SCHED_DEBUG
7273 void print_cfs_stats(struct seq_file *m, int cpu)
7275 struct cfs_rq *cfs_rq;
7278 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7279 print_cfs_rq(m, cpu, cfs_rq);
7284 __init void init_sched_fair_class(void)
7287 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7289 #ifdef CONFIG_NO_HZ_COMMON
7290 nohz.next_balance = jiffies;
7291 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7292 cpu_notifier(sched_ilb_notifier, 0);