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 inline void __update_task_entity_contrib(struct sched_entity *se);
686 /* Give new task start runnable values to heavy its load in infant time */
687 void init_task_runnable_average(struct task_struct *p)
691 p->se.avg.decay_count = 0;
692 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
693 p->se.avg.runnable_avg_sum = slice;
694 p->se.avg.runnable_avg_period = slice;
695 __update_task_entity_contrib(&p->se);
698 void init_task_runnable_average(struct task_struct *p)
704 * Update the current task's runtime statistics. Skip current tasks that
705 * are not in our scheduling class.
708 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
709 unsigned long delta_exec)
711 unsigned long delta_exec_weighted;
713 schedstat_set(curr->statistics.exec_max,
714 max((u64)delta_exec, curr->statistics.exec_max));
716 curr->sum_exec_runtime += delta_exec;
717 schedstat_add(cfs_rq, exec_clock, delta_exec);
718 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
720 curr->vruntime += delta_exec_weighted;
721 update_min_vruntime(cfs_rq);
724 static void update_curr(struct cfs_rq *cfs_rq)
726 struct sched_entity *curr = cfs_rq->curr;
727 u64 now = rq_clock_task(rq_of(cfs_rq));
728 unsigned long delta_exec;
734 * Get the amount of time the current task was running
735 * since the last time we changed load (this cannot
736 * overflow on 32 bits):
738 delta_exec = (unsigned long)(now - curr->exec_start);
742 __update_curr(cfs_rq, curr, delta_exec);
743 curr->exec_start = now;
745 if (entity_is_task(curr)) {
746 struct task_struct *curtask = task_of(curr);
748 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
749 cpuacct_charge(curtask, delta_exec);
750 account_group_exec_runtime(curtask, delta_exec);
753 account_cfs_rq_runtime(cfs_rq, delta_exec);
757 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
759 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
763 * Task is being enqueued - update stats:
765 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * Are we enqueueing a waiting task? (for current tasks
769 * a dequeue/enqueue event is a NOP)
771 if (se != cfs_rq->curr)
772 update_stats_wait_start(cfs_rq, se);
776 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
779 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
780 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
781 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
782 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
783 #ifdef CONFIG_SCHEDSTATS
784 if (entity_is_task(se)) {
785 trace_sched_stat_wait(task_of(se),
786 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
789 schedstat_set(se->statistics.wait_start, 0);
793 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
796 * Mark the end of the wait period if dequeueing a
799 if (se != cfs_rq->curr)
800 update_stats_wait_end(cfs_rq, se);
804 * We are picking a new current task - update its stats:
807 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
810 * We are starting a new run period:
812 se->exec_start = rq_clock_task(rq_of(cfs_rq));
815 /**************************************************
816 * Scheduling class queueing methods:
819 #ifdef CONFIG_NUMA_BALANCING
821 * Approximate time to scan a full NUMA task in ms. The task scan period is
822 * calculated based on the tasks virtual memory size and
823 * numa_balancing_scan_size.
825 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
826 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
827 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
829 /* Portion of address space to scan in MB */
830 unsigned int sysctl_numa_balancing_scan_size = 256;
832 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
833 unsigned int sysctl_numa_balancing_scan_delay = 1000;
835 static unsigned int task_nr_scan_windows(struct task_struct *p)
837 unsigned long rss = 0;
838 unsigned long nr_scan_pages;
841 * Calculations based on RSS as non-present and empty pages are skipped
842 * by the PTE scanner and NUMA hinting faults should be trapped based
845 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
846 rss = get_mm_rss(p->mm);
850 rss = round_up(rss, nr_scan_pages);
851 return rss / nr_scan_pages;
854 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
855 #define MAX_SCAN_WINDOW 2560
857 static unsigned int task_scan_min(struct task_struct *p)
859 unsigned int scan, floor;
860 unsigned int windows = 1;
862 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
863 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
864 floor = 1000 / windows;
866 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
867 return max_t(unsigned int, floor, scan);
870 static unsigned int task_scan_max(struct task_struct *p)
872 unsigned int smin = task_scan_min(p);
875 /* Watch for min being lower than max due to floor calculations */
876 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
877 return max(smin, smax);
880 static void task_numa_placement(struct task_struct *p)
884 if (!p->mm) /* for example, ksmd faulting in a user's mm */
886 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
887 if (p->numa_scan_seq == seq)
889 p->numa_scan_seq = seq;
890 p->numa_scan_period_max = task_scan_max(p);
892 /* FIXME: Scheduling placement policy hints go here */
896 * Got a PROT_NONE fault for a page on @node.
898 void task_numa_fault(int node, int pages, bool migrated)
900 struct task_struct *p = current;
902 if (!numabalancing_enabled)
905 /* FIXME: Allocate task-specific structure for placement policy here */
908 * If pages are properly placed (did not migrate) then scan slower.
909 * This is reset periodically in case of phase changes
912 /* Initialise if necessary */
913 if (!p->numa_scan_period_max)
914 p->numa_scan_period_max = task_scan_max(p);
916 p->numa_scan_period = min(p->numa_scan_period_max,
917 p->numa_scan_period + 10);
920 task_numa_placement(p);
923 static void reset_ptenuma_scan(struct task_struct *p)
925 ACCESS_ONCE(p->mm->numa_scan_seq)++;
926 p->mm->numa_scan_offset = 0;
930 * The expensive part of numa migration is done from task_work context.
931 * Triggered from task_tick_numa().
933 void task_numa_work(struct callback_head *work)
935 unsigned long migrate, next_scan, now = jiffies;
936 struct task_struct *p = current;
937 struct mm_struct *mm = p->mm;
938 struct vm_area_struct *vma;
939 unsigned long start, end;
940 unsigned long nr_pte_updates = 0;
943 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
945 work->next = work; /* protect against double add */
947 * Who cares about NUMA placement when they're dying.
949 * NOTE: make sure not to dereference p->mm before this check,
950 * exit_task_work() happens _after_ exit_mm() so we could be called
951 * without p->mm even though we still had it when we enqueued this
954 if (p->flags & PF_EXITING)
957 if (!mm->numa_next_reset || !mm->numa_next_scan) {
958 mm->numa_next_scan = now +
959 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
960 mm->numa_next_reset = now +
961 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
965 * Reset the scan period if enough time has gone by. Objective is that
966 * scanning will be reduced if pages are properly placed. As tasks
967 * can enter different phases this needs to be re-examined. Lacking
968 * proper tracking of reference behaviour, this blunt hammer is used.
970 migrate = mm->numa_next_reset;
971 if (time_after(now, migrate)) {
972 p->numa_scan_period = task_scan_min(p);
973 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
974 xchg(&mm->numa_next_reset, next_scan);
978 * Enforce maximal scan/migration frequency..
980 migrate = mm->numa_next_scan;
981 if (time_before(now, migrate))
984 if (p->numa_scan_period == 0) {
985 p->numa_scan_period_max = task_scan_max(p);
986 p->numa_scan_period = task_scan_min(p);
989 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
990 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
994 * Delay this task enough that another task of this mm will likely win
995 * the next time around.
997 p->node_stamp += 2 * TICK_NSEC;
999 start = mm->numa_scan_offset;
1000 pages = sysctl_numa_balancing_scan_size;
1001 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1005 down_read(&mm->mmap_sem);
1006 vma = find_vma(mm, start);
1008 reset_ptenuma_scan(p);
1012 for (; vma; vma = vma->vm_next) {
1013 if (!vma_migratable(vma))
1016 /* Skip small VMAs. They are not likely to be of relevance */
1017 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1021 start = max(start, vma->vm_start);
1022 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1023 end = min(end, vma->vm_end);
1024 nr_pte_updates += change_prot_numa(vma, start, end);
1027 * Scan sysctl_numa_balancing_scan_size but ensure that
1028 * at least one PTE is updated so that unused virtual
1029 * address space is quickly skipped.
1032 pages -= (end - start) >> PAGE_SHIFT;
1037 } while (end != vma->vm_end);
1042 * If the whole process was scanned without updates then no NUMA
1043 * hinting faults are being recorded and scan rate should be lower.
1045 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1046 p->numa_scan_period = min(p->numa_scan_period_max,
1047 p->numa_scan_period << 1);
1049 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1050 mm->numa_next_scan = next_scan;
1054 * It is possible to reach the end of the VMA list but the last few
1055 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1056 * would find the !migratable VMA on the next scan but not reset the
1057 * scanner to the start so check it now.
1060 mm->numa_scan_offset = start;
1062 reset_ptenuma_scan(p);
1063 up_read(&mm->mmap_sem);
1067 * Drive the periodic memory faults..
1069 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1071 struct callback_head *work = &curr->numa_work;
1075 * We don't care about NUMA placement if we don't have memory.
1077 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1081 * Using runtime rather than walltime has the dual advantage that
1082 * we (mostly) drive the selection from busy threads and that the
1083 * task needs to have done some actual work before we bother with
1086 now = curr->se.sum_exec_runtime;
1087 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1089 if (now - curr->node_stamp > period) {
1090 if (!curr->node_stamp)
1091 curr->numa_scan_period = task_scan_min(curr);
1092 curr->node_stamp += period;
1094 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1095 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1096 task_work_add(curr, work, true);
1101 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1104 #endif /* CONFIG_NUMA_BALANCING */
1107 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1109 update_load_add(&cfs_rq->load, se->load.weight);
1110 if (!parent_entity(se))
1111 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1113 if (entity_is_task(se))
1114 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1116 cfs_rq->nr_running++;
1120 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1122 update_load_sub(&cfs_rq->load, se->load.weight);
1123 if (!parent_entity(se))
1124 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1125 if (entity_is_task(se))
1126 list_del_init(&se->group_node);
1127 cfs_rq->nr_running--;
1130 #ifdef CONFIG_FAIR_GROUP_SCHED
1132 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1137 * Use this CPU's actual weight instead of the last load_contribution
1138 * to gain a more accurate current total weight. See
1139 * update_cfs_rq_load_contribution().
1141 tg_weight = atomic_long_read(&tg->load_avg);
1142 tg_weight -= cfs_rq->tg_load_contrib;
1143 tg_weight += cfs_rq->load.weight;
1148 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1150 long tg_weight, load, shares;
1152 tg_weight = calc_tg_weight(tg, cfs_rq);
1153 load = cfs_rq->load.weight;
1155 shares = (tg->shares * load);
1157 shares /= tg_weight;
1159 if (shares < MIN_SHARES)
1160 shares = MIN_SHARES;
1161 if (shares > tg->shares)
1162 shares = tg->shares;
1166 # else /* CONFIG_SMP */
1167 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1171 # endif /* CONFIG_SMP */
1172 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1173 unsigned long weight)
1176 /* commit outstanding execution time */
1177 if (cfs_rq->curr == se)
1178 update_curr(cfs_rq);
1179 account_entity_dequeue(cfs_rq, se);
1182 update_load_set(&se->load, weight);
1185 account_entity_enqueue(cfs_rq, se);
1188 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1190 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1192 struct task_group *tg;
1193 struct sched_entity *se;
1197 se = tg->se[cpu_of(rq_of(cfs_rq))];
1198 if (!se || throttled_hierarchy(cfs_rq))
1201 if (likely(se->load.weight == tg->shares))
1204 shares = calc_cfs_shares(cfs_rq, tg);
1206 reweight_entity(cfs_rq_of(se), se, shares);
1208 #else /* CONFIG_FAIR_GROUP_SCHED */
1209 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1212 #endif /* CONFIG_FAIR_GROUP_SCHED */
1216 * We choose a half-life close to 1 scheduling period.
1217 * Note: The tables below are dependent on this value.
1219 #define LOAD_AVG_PERIOD 32
1220 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1221 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1223 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1224 static const u32 runnable_avg_yN_inv[] = {
1225 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1226 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1227 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1228 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1229 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1230 0x85aac367, 0x82cd8698,
1234 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1235 * over-estimates when re-combining.
1237 static const u32 runnable_avg_yN_sum[] = {
1238 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1239 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1240 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1245 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1247 static __always_inline u64 decay_load(u64 val, u64 n)
1249 unsigned int local_n;
1253 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1256 /* after bounds checking we can collapse to 32-bit */
1260 * As y^PERIOD = 1/2, we can combine
1261 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1262 * With a look-up table which covers k^n (n<PERIOD)
1264 * To achieve constant time decay_load.
1266 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1267 val >>= local_n / LOAD_AVG_PERIOD;
1268 local_n %= LOAD_AVG_PERIOD;
1271 val *= runnable_avg_yN_inv[local_n];
1272 /* We don't use SRR here since we always want to round down. */
1277 * For updates fully spanning n periods, the contribution to runnable
1278 * average will be: \Sum 1024*y^n
1280 * We can compute this reasonably efficiently by combining:
1281 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1283 static u32 __compute_runnable_contrib(u64 n)
1287 if (likely(n <= LOAD_AVG_PERIOD))
1288 return runnable_avg_yN_sum[n];
1289 else if (unlikely(n >= LOAD_AVG_MAX_N))
1290 return LOAD_AVG_MAX;
1292 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1294 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1295 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1297 n -= LOAD_AVG_PERIOD;
1298 } while (n > LOAD_AVG_PERIOD);
1300 contrib = decay_load(contrib, n);
1301 return contrib + runnable_avg_yN_sum[n];
1305 * We can represent the historical contribution to runnable average as the
1306 * coefficients of a geometric series. To do this we sub-divide our runnable
1307 * history into segments of approximately 1ms (1024us); label the segment that
1308 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1310 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1312 * (now) (~1ms ago) (~2ms ago)
1314 * Let u_i denote the fraction of p_i that the entity was runnable.
1316 * We then designate the fractions u_i as our co-efficients, yielding the
1317 * following representation of historical load:
1318 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1320 * We choose y based on the with of a reasonably scheduling period, fixing:
1323 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1324 * approximately half as much as the contribution to load within the last ms
1327 * When a period "rolls over" and we have new u_0`, multiplying the previous
1328 * sum again by y is sufficient to update:
1329 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1330 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1332 static __always_inline int __update_entity_runnable_avg(u64 now,
1333 struct sched_avg *sa,
1337 u32 runnable_contrib;
1338 int delta_w, decayed = 0;
1340 delta = now - sa->last_runnable_update;
1342 * This should only happen when time goes backwards, which it
1343 * unfortunately does during sched clock init when we swap over to TSC.
1345 if ((s64)delta < 0) {
1346 sa->last_runnable_update = now;
1351 * Use 1024ns as the unit of measurement since it's a reasonable
1352 * approximation of 1us and fast to compute.
1357 sa->last_runnable_update = now;
1359 /* delta_w is the amount already accumulated against our next period */
1360 delta_w = sa->runnable_avg_period % 1024;
1361 if (delta + delta_w >= 1024) {
1362 /* period roll-over */
1366 * Now that we know we're crossing a period boundary, figure
1367 * out how much from delta we need to complete the current
1368 * period and accrue it.
1370 delta_w = 1024 - delta_w;
1372 sa->runnable_avg_sum += delta_w;
1373 sa->runnable_avg_period += delta_w;
1377 /* Figure out how many additional periods this update spans */
1378 periods = delta / 1024;
1381 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1383 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1386 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1387 runnable_contrib = __compute_runnable_contrib(periods);
1389 sa->runnable_avg_sum += runnable_contrib;
1390 sa->runnable_avg_period += runnable_contrib;
1393 /* Remainder of delta accrued against u_0` */
1395 sa->runnable_avg_sum += delta;
1396 sa->runnable_avg_period += delta;
1401 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1402 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1404 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1405 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1407 decays -= se->avg.decay_count;
1411 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1412 se->avg.decay_count = 0;
1417 #ifdef CONFIG_FAIR_GROUP_SCHED
1418 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1421 struct task_group *tg = cfs_rq->tg;
1424 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1425 tg_contrib -= cfs_rq->tg_load_contrib;
1427 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1428 atomic_long_add(tg_contrib, &tg->load_avg);
1429 cfs_rq->tg_load_contrib += tg_contrib;
1434 * Aggregate cfs_rq runnable averages into an equivalent task_group
1435 * representation for computing load contributions.
1437 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1438 struct cfs_rq *cfs_rq)
1440 struct task_group *tg = cfs_rq->tg;
1443 /* The fraction of a cpu used by this cfs_rq */
1444 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1445 sa->runnable_avg_period + 1);
1446 contrib -= cfs_rq->tg_runnable_contrib;
1448 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1449 atomic_add(contrib, &tg->runnable_avg);
1450 cfs_rq->tg_runnable_contrib += contrib;
1454 static inline void __update_group_entity_contrib(struct sched_entity *se)
1456 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1457 struct task_group *tg = cfs_rq->tg;
1462 contrib = cfs_rq->tg_load_contrib * tg->shares;
1463 se->avg.load_avg_contrib = div_u64(contrib,
1464 atomic_long_read(&tg->load_avg) + 1);
1467 * For group entities we need to compute a correction term in the case
1468 * that they are consuming <1 cpu so that we would contribute the same
1469 * load as a task of equal weight.
1471 * Explicitly co-ordinating this measurement would be expensive, but
1472 * fortunately the sum of each cpus contribution forms a usable
1473 * lower-bound on the true value.
1475 * Consider the aggregate of 2 contributions. Either they are disjoint
1476 * (and the sum represents true value) or they are disjoint and we are
1477 * understating by the aggregate of their overlap.
1479 * Extending this to N cpus, for a given overlap, the maximum amount we
1480 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1481 * cpus that overlap for this interval and w_i is the interval width.
1483 * On a small machine; the first term is well-bounded which bounds the
1484 * total error since w_i is a subset of the period. Whereas on a
1485 * larger machine, while this first term can be larger, if w_i is the
1486 * of consequential size guaranteed to see n_i*w_i quickly converge to
1487 * our upper bound of 1-cpu.
1489 runnable_avg = atomic_read(&tg->runnable_avg);
1490 if (runnable_avg < NICE_0_LOAD) {
1491 se->avg.load_avg_contrib *= runnable_avg;
1492 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1496 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1497 int force_update) {}
1498 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1499 struct cfs_rq *cfs_rq) {}
1500 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1503 static inline void __update_task_entity_contrib(struct sched_entity *se)
1507 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1508 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1509 contrib /= (se->avg.runnable_avg_period + 1);
1510 se->avg.load_avg_contrib = scale_load(contrib);
1513 /* Compute the current contribution to load_avg by se, return any delta */
1514 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1516 long old_contrib = se->avg.load_avg_contrib;
1518 if (entity_is_task(se)) {
1519 __update_task_entity_contrib(se);
1521 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1522 __update_group_entity_contrib(se);
1525 return se->avg.load_avg_contrib - old_contrib;
1528 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1531 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1532 cfs_rq->blocked_load_avg -= load_contrib;
1534 cfs_rq->blocked_load_avg = 0;
1537 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1539 /* Update a sched_entity's runnable average */
1540 static inline void update_entity_load_avg(struct sched_entity *se,
1543 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1548 * For a group entity we need to use their owned cfs_rq_clock_task() in
1549 * case they are the parent of a throttled hierarchy.
1551 if (entity_is_task(se))
1552 now = cfs_rq_clock_task(cfs_rq);
1554 now = cfs_rq_clock_task(group_cfs_rq(se));
1556 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1559 contrib_delta = __update_entity_load_avg_contrib(se);
1565 cfs_rq->runnable_load_avg += contrib_delta;
1567 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1571 * Decay the load contributed by all blocked children and account this so that
1572 * their contribution may appropriately discounted when they wake up.
1574 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1576 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1579 decays = now - cfs_rq->last_decay;
1580 if (!decays && !force_update)
1583 if (atomic_long_read(&cfs_rq->removed_load)) {
1584 unsigned long removed_load;
1585 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1586 subtract_blocked_load_contrib(cfs_rq, removed_load);
1590 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1592 atomic64_add(decays, &cfs_rq->decay_counter);
1593 cfs_rq->last_decay = now;
1596 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1599 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1601 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1602 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1605 /* Add the load generated by se into cfs_rq's child load-average */
1606 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1607 struct sched_entity *se,
1611 * We track migrations using entity decay_count <= 0, on a wake-up
1612 * migration we use a negative decay count to track the remote decays
1613 * accumulated while sleeping.
1615 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1616 * are seen by enqueue_entity_load_avg() as a migration with an already
1617 * constructed load_avg_contrib.
1619 if (unlikely(se->avg.decay_count <= 0)) {
1620 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1621 if (se->avg.decay_count) {
1623 * In a wake-up migration we have to approximate the
1624 * time sleeping. This is because we can't synchronize
1625 * clock_task between the two cpus, and it is not
1626 * guaranteed to be read-safe. Instead, we can
1627 * approximate this using our carried decays, which are
1628 * explicitly atomically readable.
1630 se->avg.last_runnable_update -= (-se->avg.decay_count)
1632 update_entity_load_avg(se, 0);
1633 /* Indicate that we're now synchronized and on-rq */
1634 se->avg.decay_count = 0;
1639 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1640 * would have made count negative); we must be careful to avoid
1641 * double-accounting blocked time after synchronizing decays.
1643 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1647 /* migrated tasks did not contribute to our blocked load */
1649 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1650 update_entity_load_avg(se, 0);
1653 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1654 /* we force update consideration on load-balancer moves */
1655 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1659 * Remove se's load from this cfs_rq child load-average, if the entity is
1660 * transitioning to a blocked state we track its projected decay using
1663 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1664 struct sched_entity *se,
1667 update_entity_load_avg(se, 1);
1668 /* we force update consideration on load-balancer moves */
1669 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1671 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1673 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1674 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1675 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1679 * Update the rq's load with the elapsed running time before entering
1680 * idle. if the last scheduled task is not a CFS task, idle_enter will
1681 * be the only way to update the runnable statistic.
1683 void idle_enter_fair(struct rq *this_rq)
1685 update_rq_runnable_avg(this_rq, 1);
1689 * Update the rq's load with the elapsed idle time before a task is
1690 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1691 * be the only way to update the runnable statistic.
1693 void idle_exit_fair(struct rq *this_rq)
1695 update_rq_runnable_avg(this_rq, 0);
1699 static inline void update_entity_load_avg(struct sched_entity *se,
1700 int update_cfs_rq) {}
1701 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1702 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1703 struct sched_entity *se,
1705 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1706 struct sched_entity *se,
1708 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1709 int force_update) {}
1712 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1714 #ifdef CONFIG_SCHEDSTATS
1715 struct task_struct *tsk = NULL;
1717 if (entity_is_task(se))
1720 if (se->statistics.sleep_start) {
1721 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1726 if (unlikely(delta > se->statistics.sleep_max))
1727 se->statistics.sleep_max = delta;
1729 se->statistics.sleep_start = 0;
1730 se->statistics.sum_sleep_runtime += delta;
1733 account_scheduler_latency(tsk, delta >> 10, 1);
1734 trace_sched_stat_sleep(tsk, delta);
1737 if (se->statistics.block_start) {
1738 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1743 if (unlikely(delta > se->statistics.block_max))
1744 se->statistics.block_max = delta;
1746 se->statistics.block_start = 0;
1747 se->statistics.sum_sleep_runtime += delta;
1750 if (tsk->in_iowait) {
1751 se->statistics.iowait_sum += delta;
1752 se->statistics.iowait_count++;
1753 trace_sched_stat_iowait(tsk, delta);
1756 trace_sched_stat_blocked(tsk, delta);
1759 * Blocking time is in units of nanosecs, so shift by
1760 * 20 to get a milliseconds-range estimation of the
1761 * amount of time that the task spent sleeping:
1763 if (unlikely(prof_on == SLEEP_PROFILING)) {
1764 profile_hits(SLEEP_PROFILING,
1765 (void *)get_wchan(tsk),
1768 account_scheduler_latency(tsk, delta >> 10, 0);
1774 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1776 #ifdef CONFIG_SCHED_DEBUG
1777 s64 d = se->vruntime - cfs_rq->min_vruntime;
1782 if (d > 3*sysctl_sched_latency)
1783 schedstat_inc(cfs_rq, nr_spread_over);
1788 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1790 u64 vruntime = cfs_rq->min_vruntime;
1793 * The 'current' period is already promised to the current tasks,
1794 * however the extra weight of the new task will slow them down a
1795 * little, place the new task so that it fits in the slot that
1796 * stays open at the end.
1798 if (initial && sched_feat(START_DEBIT))
1799 vruntime += sched_vslice(cfs_rq, se);
1801 /* sleeps up to a single latency don't count. */
1803 unsigned long thresh = sysctl_sched_latency;
1806 * Halve their sleep time's effect, to allow
1807 * for a gentler effect of sleepers:
1809 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1815 /* ensure we never gain time by being placed backwards. */
1816 se->vruntime = max_vruntime(se->vruntime, vruntime);
1819 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1822 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1825 * Update the normalized vruntime before updating min_vruntime
1826 * through calling update_curr().
1828 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1829 se->vruntime += cfs_rq->min_vruntime;
1832 * Update run-time statistics of the 'current'.
1834 update_curr(cfs_rq);
1835 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1836 account_entity_enqueue(cfs_rq, se);
1837 update_cfs_shares(cfs_rq);
1839 if (flags & ENQUEUE_WAKEUP) {
1840 place_entity(cfs_rq, se, 0);
1841 enqueue_sleeper(cfs_rq, se);
1844 update_stats_enqueue(cfs_rq, se);
1845 check_spread(cfs_rq, se);
1846 if (se != cfs_rq->curr)
1847 __enqueue_entity(cfs_rq, se);
1850 if (cfs_rq->nr_running == 1) {
1851 list_add_leaf_cfs_rq(cfs_rq);
1852 check_enqueue_throttle(cfs_rq);
1856 static void __clear_buddies_last(struct sched_entity *se)
1858 for_each_sched_entity(se) {
1859 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1860 if (cfs_rq->last == se)
1861 cfs_rq->last = NULL;
1867 static void __clear_buddies_next(struct sched_entity *se)
1869 for_each_sched_entity(se) {
1870 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1871 if (cfs_rq->next == se)
1872 cfs_rq->next = NULL;
1878 static void __clear_buddies_skip(struct sched_entity *se)
1880 for_each_sched_entity(se) {
1881 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1882 if (cfs_rq->skip == se)
1883 cfs_rq->skip = NULL;
1889 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1891 if (cfs_rq->last == se)
1892 __clear_buddies_last(se);
1894 if (cfs_rq->next == se)
1895 __clear_buddies_next(se);
1897 if (cfs_rq->skip == se)
1898 __clear_buddies_skip(se);
1901 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1904 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1907 * Update run-time statistics of the 'current'.
1909 update_curr(cfs_rq);
1910 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1912 update_stats_dequeue(cfs_rq, se);
1913 if (flags & DEQUEUE_SLEEP) {
1914 #ifdef CONFIG_SCHEDSTATS
1915 if (entity_is_task(se)) {
1916 struct task_struct *tsk = task_of(se);
1918 if (tsk->state & TASK_INTERRUPTIBLE)
1919 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1920 if (tsk->state & TASK_UNINTERRUPTIBLE)
1921 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1926 clear_buddies(cfs_rq, se);
1928 if (se != cfs_rq->curr)
1929 __dequeue_entity(cfs_rq, se);
1931 account_entity_dequeue(cfs_rq, se);
1934 * Normalize the entity after updating the min_vruntime because the
1935 * update can refer to the ->curr item and we need to reflect this
1936 * movement in our normalized position.
1938 if (!(flags & DEQUEUE_SLEEP))
1939 se->vruntime -= cfs_rq->min_vruntime;
1941 /* return excess runtime on last dequeue */
1942 return_cfs_rq_runtime(cfs_rq);
1944 update_min_vruntime(cfs_rq);
1945 update_cfs_shares(cfs_rq);
1949 * Preempt the current task with a newly woken task if needed:
1952 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1954 unsigned long ideal_runtime, delta_exec;
1955 struct sched_entity *se;
1958 ideal_runtime = sched_slice(cfs_rq, curr);
1959 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1960 if (delta_exec > ideal_runtime) {
1961 resched_task(rq_of(cfs_rq)->curr);
1963 * The current task ran long enough, ensure it doesn't get
1964 * re-elected due to buddy favours.
1966 clear_buddies(cfs_rq, curr);
1971 * Ensure that a task that missed wakeup preemption by a
1972 * narrow margin doesn't have to wait for a full slice.
1973 * This also mitigates buddy induced latencies under load.
1975 if (delta_exec < sysctl_sched_min_granularity)
1978 se = __pick_first_entity(cfs_rq);
1979 delta = curr->vruntime - se->vruntime;
1984 if (delta > ideal_runtime)
1985 resched_task(rq_of(cfs_rq)->curr);
1989 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1991 /* 'current' is not kept within the tree. */
1994 * Any task has to be enqueued before it get to execute on
1995 * a CPU. So account for the time it spent waiting on the
1998 update_stats_wait_end(cfs_rq, se);
1999 __dequeue_entity(cfs_rq, se);
2002 update_stats_curr_start(cfs_rq, se);
2004 #ifdef CONFIG_SCHEDSTATS
2006 * Track our maximum slice length, if the CPU's load is at
2007 * least twice that of our own weight (i.e. dont track it
2008 * when there are only lesser-weight tasks around):
2010 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2011 se->statistics.slice_max = max(se->statistics.slice_max,
2012 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2015 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2019 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2022 * Pick the next process, keeping these things in mind, in this order:
2023 * 1) keep things fair between processes/task groups
2024 * 2) pick the "next" process, since someone really wants that to run
2025 * 3) pick the "last" process, for cache locality
2026 * 4) do not run the "skip" process, if something else is available
2028 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2030 struct sched_entity *se = __pick_first_entity(cfs_rq);
2031 struct sched_entity *left = se;
2034 * Avoid running the skip buddy, if running something else can
2035 * be done without getting too unfair.
2037 if (cfs_rq->skip == se) {
2038 struct sched_entity *second = __pick_next_entity(se);
2039 if (second && wakeup_preempt_entity(second, left) < 1)
2044 * Prefer last buddy, try to return the CPU to a preempted task.
2046 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2050 * Someone really wants this to run. If it's not unfair, run it.
2052 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2055 clear_buddies(cfs_rq, se);
2060 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2062 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2065 * If still on the runqueue then deactivate_task()
2066 * was not called and update_curr() has to be done:
2069 update_curr(cfs_rq);
2071 /* throttle cfs_rqs exceeding runtime */
2072 check_cfs_rq_runtime(cfs_rq);
2074 check_spread(cfs_rq, prev);
2076 update_stats_wait_start(cfs_rq, prev);
2077 /* Put 'current' back into the tree. */
2078 __enqueue_entity(cfs_rq, prev);
2079 /* in !on_rq case, update occurred at dequeue */
2080 update_entity_load_avg(prev, 1);
2082 cfs_rq->curr = NULL;
2086 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2089 * Update run-time statistics of the 'current'.
2091 update_curr(cfs_rq);
2094 * Ensure that runnable average is periodically updated.
2096 update_entity_load_avg(curr, 1);
2097 update_cfs_rq_blocked_load(cfs_rq, 1);
2098 update_cfs_shares(cfs_rq);
2100 #ifdef CONFIG_SCHED_HRTICK
2102 * queued ticks are scheduled to match the slice, so don't bother
2103 * validating it and just reschedule.
2106 resched_task(rq_of(cfs_rq)->curr);
2110 * don't let the period tick interfere with the hrtick preemption
2112 if (!sched_feat(DOUBLE_TICK) &&
2113 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2117 if (cfs_rq->nr_running > 1)
2118 check_preempt_tick(cfs_rq, curr);
2122 /**************************************************
2123 * CFS bandwidth control machinery
2126 #ifdef CONFIG_CFS_BANDWIDTH
2128 #ifdef HAVE_JUMP_LABEL
2129 static struct static_key __cfs_bandwidth_used;
2131 static inline bool cfs_bandwidth_used(void)
2133 return static_key_false(&__cfs_bandwidth_used);
2136 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2138 /* only need to count groups transitioning between enabled/!enabled */
2139 if (enabled && !was_enabled)
2140 static_key_slow_inc(&__cfs_bandwidth_used);
2141 else if (!enabled && was_enabled)
2142 static_key_slow_dec(&__cfs_bandwidth_used);
2144 #else /* HAVE_JUMP_LABEL */
2145 static bool cfs_bandwidth_used(void)
2150 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2151 #endif /* HAVE_JUMP_LABEL */
2154 * default period for cfs group bandwidth.
2155 * default: 0.1s, units: nanoseconds
2157 static inline u64 default_cfs_period(void)
2159 return 100000000ULL;
2162 static inline u64 sched_cfs_bandwidth_slice(void)
2164 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2168 * Replenish runtime according to assigned quota and update expiration time.
2169 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2170 * additional synchronization around rq->lock.
2172 * requires cfs_b->lock
2174 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2178 if (cfs_b->quota == RUNTIME_INF)
2181 now = sched_clock_cpu(smp_processor_id());
2182 cfs_b->runtime = cfs_b->quota;
2183 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2186 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2188 return &tg->cfs_bandwidth;
2191 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2192 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2194 if (unlikely(cfs_rq->throttle_count))
2195 return cfs_rq->throttled_clock_task;
2197 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2200 /* returns 0 on failure to allocate runtime */
2201 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2203 struct task_group *tg = cfs_rq->tg;
2204 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2205 u64 amount = 0, min_amount, expires;
2207 /* note: this is a positive sum as runtime_remaining <= 0 */
2208 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2210 raw_spin_lock(&cfs_b->lock);
2211 if (cfs_b->quota == RUNTIME_INF)
2212 amount = min_amount;
2215 * If the bandwidth pool has become inactive, then at least one
2216 * period must have elapsed since the last consumption.
2217 * Refresh the global state and ensure bandwidth timer becomes
2220 if (!cfs_b->timer_active) {
2221 __refill_cfs_bandwidth_runtime(cfs_b);
2222 __start_cfs_bandwidth(cfs_b);
2225 if (cfs_b->runtime > 0) {
2226 amount = min(cfs_b->runtime, min_amount);
2227 cfs_b->runtime -= amount;
2231 expires = cfs_b->runtime_expires;
2232 raw_spin_unlock(&cfs_b->lock);
2234 cfs_rq->runtime_remaining += amount;
2236 * we may have advanced our local expiration to account for allowed
2237 * spread between our sched_clock and the one on which runtime was
2240 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2241 cfs_rq->runtime_expires = expires;
2243 return cfs_rq->runtime_remaining > 0;
2247 * Note: This depends on the synchronization provided by sched_clock and the
2248 * fact that rq->clock snapshots this value.
2250 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2252 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2254 /* if the deadline is ahead of our clock, nothing to do */
2255 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2258 if (cfs_rq->runtime_remaining < 0)
2262 * If the local deadline has passed we have to consider the
2263 * possibility that our sched_clock is 'fast' and the global deadline
2264 * has not truly expired.
2266 * Fortunately we can check determine whether this the case by checking
2267 * whether the global deadline has advanced.
2270 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2271 /* extend local deadline, drift is bounded above by 2 ticks */
2272 cfs_rq->runtime_expires += TICK_NSEC;
2274 /* global deadline is ahead, expiration has passed */
2275 cfs_rq->runtime_remaining = 0;
2279 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2280 unsigned long delta_exec)
2282 /* dock delta_exec before expiring quota (as it could span periods) */
2283 cfs_rq->runtime_remaining -= delta_exec;
2284 expire_cfs_rq_runtime(cfs_rq);
2286 if (likely(cfs_rq->runtime_remaining > 0))
2290 * if we're unable to extend our runtime we resched so that the active
2291 * hierarchy can be throttled
2293 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2294 resched_task(rq_of(cfs_rq)->curr);
2297 static __always_inline
2298 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2300 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2303 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2306 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2308 return cfs_bandwidth_used() && cfs_rq->throttled;
2311 /* check whether cfs_rq, or any parent, is throttled */
2312 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2314 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2318 * Ensure that neither of the group entities corresponding to src_cpu or
2319 * dest_cpu are members of a throttled hierarchy when performing group
2320 * load-balance operations.
2322 static inline int throttled_lb_pair(struct task_group *tg,
2323 int src_cpu, int dest_cpu)
2325 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2327 src_cfs_rq = tg->cfs_rq[src_cpu];
2328 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2330 return throttled_hierarchy(src_cfs_rq) ||
2331 throttled_hierarchy(dest_cfs_rq);
2334 /* updated child weight may affect parent so we have to do this bottom up */
2335 static int tg_unthrottle_up(struct task_group *tg, void *data)
2337 struct rq *rq = data;
2338 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2340 cfs_rq->throttle_count--;
2342 if (!cfs_rq->throttle_count) {
2343 /* adjust cfs_rq_clock_task() */
2344 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2345 cfs_rq->throttled_clock_task;
2352 static int tg_throttle_down(struct task_group *tg, void *data)
2354 struct rq *rq = data;
2355 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2357 /* group is entering throttled state, stop time */
2358 if (!cfs_rq->throttle_count)
2359 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2360 cfs_rq->throttle_count++;
2365 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2367 struct rq *rq = rq_of(cfs_rq);
2368 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2369 struct sched_entity *se;
2370 long task_delta, dequeue = 1;
2372 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2374 /* freeze hierarchy runnable averages while throttled */
2376 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2379 task_delta = cfs_rq->h_nr_running;
2380 for_each_sched_entity(se) {
2381 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2382 /* throttled entity or throttle-on-deactivate */
2387 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2388 qcfs_rq->h_nr_running -= task_delta;
2390 if (qcfs_rq->load.weight)
2395 rq->nr_running -= task_delta;
2397 cfs_rq->throttled = 1;
2398 cfs_rq->throttled_clock = rq_clock(rq);
2399 raw_spin_lock(&cfs_b->lock);
2400 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2401 raw_spin_unlock(&cfs_b->lock);
2404 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2406 struct rq *rq = rq_of(cfs_rq);
2407 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2408 struct sched_entity *se;
2412 se = cfs_rq->tg->se[cpu_of(rq)];
2414 cfs_rq->throttled = 0;
2416 update_rq_clock(rq);
2418 raw_spin_lock(&cfs_b->lock);
2419 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2420 list_del_rcu(&cfs_rq->throttled_list);
2421 raw_spin_unlock(&cfs_b->lock);
2423 /* update hierarchical throttle state */
2424 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2426 if (!cfs_rq->load.weight)
2429 task_delta = cfs_rq->h_nr_running;
2430 for_each_sched_entity(se) {
2434 cfs_rq = cfs_rq_of(se);
2436 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2437 cfs_rq->h_nr_running += task_delta;
2439 if (cfs_rq_throttled(cfs_rq))
2444 rq->nr_running += task_delta;
2446 /* determine whether we need to wake up potentially idle cpu */
2447 if (rq->curr == rq->idle && rq->cfs.nr_running)
2448 resched_task(rq->curr);
2451 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2452 u64 remaining, u64 expires)
2454 struct cfs_rq *cfs_rq;
2455 u64 runtime = remaining;
2458 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2460 struct rq *rq = rq_of(cfs_rq);
2462 raw_spin_lock(&rq->lock);
2463 if (!cfs_rq_throttled(cfs_rq))
2466 runtime = -cfs_rq->runtime_remaining + 1;
2467 if (runtime > remaining)
2468 runtime = remaining;
2469 remaining -= runtime;
2471 cfs_rq->runtime_remaining += runtime;
2472 cfs_rq->runtime_expires = expires;
2474 /* we check whether we're throttled above */
2475 if (cfs_rq->runtime_remaining > 0)
2476 unthrottle_cfs_rq(cfs_rq);
2479 raw_spin_unlock(&rq->lock);
2490 * Responsible for refilling a task_group's bandwidth and unthrottling its
2491 * cfs_rqs as appropriate. If there has been no activity within the last
2492 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2493 * used to track this state.
2495 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2497 u64 runtime, runtime_expires;
2498 int idle = 1, throttled;
2500 raw_spin_lock(&cfs_b->lock);
2501 /* no need to continue the timer with no bandwidth constraint */
2502 if (cfs_b->quota == RUNTIME_INF)
2505 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2506 /* idle depends on !throttled (for the case of a large deficit) */
2507 idle = cfs_b->idle && !throttled;
2508 cfs_b->nr_periods += overrun;
2510 /* if we're going inactive then everything else can be deferred */
2514 __refill_cfs_bandwidth_runtime(cfs_b);
2517 /* mark as potentially idle for the upcoming period */
2522 /* account preceding periods in which throttling occurred */
2523 cfs_b->nr_throttled += overrun;
2526 * There are throttled entities so we must first use the new bandwidth
2527 * to unthrottle them before making it generally available. This
2528 * ensures that all existing debts will be paid before a new cfs_rq is
2531 runtime = cfs_b->runtime;
2532 runtime_expires = cfs_b->runtime_expires;
2536 * This check is repeated as we are holding onto the new bandwidth
2537 * while we unthrottle. This can potentially race with an unthrottled
2538 * group trying to acquire new bandwidth from the global pool.
2540 while (throttled && runtime > 0) {
2541 raw_spin_unlock(&cfs_b->lock);
2542 /* we can't nest cfs_b->lock while distributing bandwidth */
2543 runtime = distribute_cfs_runtime(cfs_b, runtime,
2545 raw_spin_lock(&cfs_b->lock);
2547 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2550 /* return (any) remaining runtime */
2551 cfs_b->runtime = runtime;
2553 * While we are ensured activity in the period following an
2554 * unthrottle, this also covers the case in which the new bandwidth is
2555 * insufficient to cover the existing bandwidth deficit. (Forcing the
2556 * timer to remain active while there are any throttled entities.)
2561 cfs_b->timer_active = 0;
2562 raw_spin_unlock(&cfs_b->lock);
2567 /* a cfs_rq won't donate quota below this amount */
2568 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2569 /* minimum remaining period time to redistribute slack quota */
2570 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2571 /* how long we wait to gather additional slack before distributing */
2572 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2574 /* are we near the end of the current quota period? */
2575 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2577 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2580 /* if the call-back is running a quota refresh is already occurring */
2581 if (hrtimer_callback_running(refresh_timer))
2584 /* is a quota refresh about to occur? */
2585 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2586 if (remaining < min_expire)
2592 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2594 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2596 /* if there's a quota refresh soon don't bother with slack */
2597 if (runtime_refresh_within(cfs_b, min_left))
2600 start_bandwidth_timer(&cfs_b->slack_timer,
2601 ns_to_ktime(cfs_bandwidth_slack_period));
2604 /* we know any runtime found here is valid as update_curr() precedes return */
2605 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2607 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2608 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2610 if (slack_runtime <= 0)
2613 raw_spin_lock(&cfs_b->lock);
2614 if (cfs_b->quota != RUNTIME_INF &&
2615 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2616 cfs_b->runtime += slack_runtime;
2618 /* we are under rq->lock, defer unthrottling using a timer */
2619 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2620 !list_empty(&cfs_b->throttled_cfs_rq))
2621 start_cfs_slack_bandwidth(cfs_b);
2623 raw_spin_unlock(&cfs_b->lock);
2625 /* even if it's not valid for return we don't want to try again */
2626 cfs_rq->runtime_remaining -= slack_runtime;
2629 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2631 if (!cfs_bandwidth_used())
2634 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2637 __return_cfs_rq_runtime(cfs_rq);
2641 * This is done with a timer (instead of inline with bandwidth return) since
2642 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2644 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2646 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2649 /* confirm we're still not at a refresh boundary */
2650 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2653 raw_spin_lock(&cfs_b->lock);
2654 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2655 runtime = cfs_b->runtime;
2658 expires = cfs_b->runtime_expires;
2659 raw_spin_unlock(&cfs_b->lock);
2664 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2666 raw_spin_lock(&cfs_b->lock);
2667 if (expires == cfs_b->runtime_expires)
2668 cfs_b->runtime = runtime;
2669 raw_spin_unlock(&cfs_b->lock);
2673 * When a group wakes up we want to make sure that its quota is not already
2674 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2675 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2677 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2679 if (!cfs_bandwidth_used())
2682 /* an active group must be handled by the update_curr()->put() path */
2683 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2686 /* ensure the group is not already throttled */
2687 if (cfs_rq_throttled(cfs_rq))
2690 /* update runtime allocation */
2691 account_cfs_rq_runtime(cfs_rq, 0);
2692 if (cfs_rq->runtime_remaining <= 0)
2693 throttle_cfs_rq(cfs_rq);
2696 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2697 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2699 if (!cfs_bandwidth_used())
2702 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2706 * it's possible for a throttled entity to be forced into a running
2707 * state (e.g. set_curr_task), in this case we're finished.
2709 if (cfs_rq_throttled(cfs_rq))
2712 throttle_cfs_rq(cfs_rq);
2715 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2717 struct cfs_bandwidth *cfs_b =
2718 container_of(timer, struct cfs_bandwidth, slack_timer);
2719 do_sched_cfs_slack_timer(cfs_b);
2721 return HRTIMER_NORESTART;
2724 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2726 struct cfs_bandwidth *cfs_b =
2727 container_of(timer, struct cfs_bandwidth, period_timer);
2733 now = hrtimer_cb_get_time(timer);
2734 overrun = hrtimer_forward(timer, now, cfs_b->period);
2739 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2742 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2745 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2747 raw_spin_lock_init(&cfs_b->lock);
2749 cfs_b->quota = RUNTIME_INF;
2750 cfs_b->period = ns_to_ktime(default_cfs_period());
2752 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2753 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2754 cfs_b->period_timer.function = sched_cfs_period_timer;
2755 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2756 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2759 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2761 cfs_rq->runtime_enabled = 0;
2762 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2765 /* requires cfs_b->lock, may release to reprogram timer */
2766 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2769 * The timer may be active because we're trying to set a new bandwidth
2770 * period or because we're racing with the tear-down path
2771 * (timer_active==0 becomes visible before the hrtimer call-back
2772 * terminates). In either case we ensure that it's re-programmed
2774 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2775 raw_spin_unlock(&cfs_b->lock);
2776 /* ensure cfs_b->lock is available while we wait */
2777 hrtimer_cancel(&cfs_b->period_timer);
2779 raw_spin_lock(&cfs_b->lock);
2780 /* if someone else restarted the timer then we're done */
2781 if (cfs_b->timer_active)
2785 cfs_b->timer_active = 1;
2786 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2789 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2791 hrtimer_cancel(&cfs_b->period_timer);
2792 hrtimer_cancel(&cfs_b->slack_timer);
2795 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2797 struct cfs_rq *cfs_rq;
2799 for_each_leaf_cfs_rq(rq, cfs_rq) {
2800 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2802 if (!cfs_rq->runtime_enabled)
2806 * clock_task is not advancing so we just need to make sure
2807 * there's some valid quota amount
2809 cfs_rq->runtime_remaining = cfs_b->quota;
2810 if (cfs_rq_throttled(cfs_rq))
2811 unthrottle_cfs_rq(cfs_rq);
2815 #else /* CONFIG_CFS_BANDWIDTH */
2816 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2818 return rq_clock_task(rq_of(cfs_rq));
2821 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2822 unsigned long delta_exec) {}
2823 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2824 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2825 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2827 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2832 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2837 static inline int throttled_lb_pair(struct task_group *tg,
2838 int src_cpu, int dest_cpu)
2843 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2845 #ifdef CONFIG_FAIR_GROUP_SCHED
2846 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2849 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2853 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2854 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2856 #endif /* CONFIG_CFS_BANDWIDTH */
2858 /**************************************************
2859 * CFS operations on tasks:
2862 #ifdef CONFIG_SCHED_HRTICK
2863 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2865 struct sched_entity *se = &p->se;
2866 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2868 WARN_ON(task_rq(p) != rq);
2870 if (cfs_rq->nr_running > 1) {
2871 u64 slice = sched_slice(cfs_rq, se);
2872 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2873 s64 delta = slice - ran;
2882 * Don't schedule slices shorter than 10000ns, that just
2883 * doesn't make sense. Rely on vruntime for fairness.
2886 delta = max_t(s64, 10000LL, delta);
2888 hrtick_start(rq, delta);
2893 * called from enqueue/dequeue and updates the hrtick when the
2894 * current task is from our class and nr_running is low enough
2897 static void hrtick_update(struct rq *rq)
2899 struct task_struct *curr = rq->curr;
2901 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2904 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2905 hrtick_start_fair(rq, curr);
2907 #else /* !CONFIG_SCHED_HRTICK */
2909 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2913 static inline void hrtick_update(struct rq *rq)
2919 * The enqueue_task method is called before nr_running is
2920 * increased. Here we update the fair scheduling stats and
2921 * then put the task into the rbtree:
2924 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2926 struct cfs_rq *cfs_rq;
2927 struct sched_entity *se = &p->se;
2929 for_each_sched_entity(se) {
2932 cfs_rq = cfs_rq_of(se);
2933 enqueue_entity(cfs_rq, se, flags);
2936 * end evaluation on encountering a throttled cfs_rq
2938 * note: in the case of encountering a throttled cfs_rq we will
2939 * post the final h_nr_running increment below.
2941 if (cfs_rq_throttled(cfs_rq))
2943 cfs_rq->h_nr_running++;
2945 flags = ENQUEUE_WAKEUP;
2948 for_each_sched_entity(se) {
2949 cfs_rq = cfs_rq_of(se);
2950 cfs_rq->h_nr_running++;
2952 if (cfs_rq_throttled(cfs_rq))
2955 update_cfs_shares(cfs_rq);
2956 update_entity_load_avg(se, 1);
2960 update_rq_runnable_avg(rq, rq->nr_running);
2966 static void set_next_buddy(struct sched_entity *se);
2969 * The dequeue_task method is called before nr_running is
2970 * decreased. We remove the task from the rbtree and
2971 * update the fair scheduling stats:
2973 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2975 struct cfs_rq *cfs_rq;
2976 struct sched_entity *se = &p->se;
2977 int task_sleep = flags & DEQUEUE_SLEEP;
2979 for_each_sched_entity(se) {
2980 cfs_rq = cfs_rq_of(se);
2981 dequeue_entity(cfs_rq, se, flags);
2984 * end evaluation on encountering a throttled cfs_rq
2986 * note: in the case of encountering a throttled cfs_rq we will
2987 * post the final h_nr_running decrement below.
2989 if (cfs_rq_throttled(cfs_rq))
2991 cfs_rq->h_nr_running--;
2993 /* Don't dequeue parent if it has other entities besides us */
2994 if (cfs_rq->load.weight) {
2996 * Bias pick_next to pick a task from this cfs_rq, as
2997 * p is sleeping when it is within its sched_slice.
2999 if (task_sleep && parent_entity(se))
3000 set_next_buddy(parent_entity(se));
3002 /* avoid re-evaluating load for this entity */
3003 se = parent_entity(se);
3006 flags |= DEQUEUE_SLEEP;
3009 for_each_sched_entity(se) {
3010 cfs_rq = cfs_rq_of(se);
3011 cfs_rq->h_nr_running--;
3013 if (cfs_rq_throttled(cfs_rq))
3016 update_cfs_shares(cfs_rq);
3017 update_entity_load_avg(se, 1);
3022 update_rq_runnable_avg(rq, 1);
3028 /* Used instead of source_load when we know the type == 0 */
3029 static unsigned long weighted_cpuload(const int cpu)
3031 return cpu_rq(cpu)->cfs.runnable_load_avg;
3035 * Return a low guess at the load of a migration-source cpu weighted
3036 * according to the scheduling class and "nice" value.
3038 * We want to under-estimate the load of migration sources, to
3039 * balance conservatively.
3041 static unsigned long source_load(int cpu, int type)
3043 struct rq *rq = cpu_rq(cpu);
3044 unsigned long total = weighted_cpuload(cpu);
3046 if (type == 0 || !sched_feat(LB_BIAS))
3049 return min(rq->cpu_load[type-1], total);
3053 * Return a high guess at the load of a migration-target cpu weighted
3054 * according to the scheduling class and "nice" value.
3056 static unsigned long target_load(int cpu, int type)
3058 struct rq *rq = cpu_rq(cpu);
3059 unsigned long total = weighted_cpuload(cpu);
3061 if (type == 0 || !sched_feat(LB_BIAS))
3064 return max(rq->cpu_load[type-1], total);
3067 static unsigned long power_of(int cpu)
3069 return cpu_rq(cpu)->cpu_power;
3072 static unsigned long cpu_avg_load_per_task(int cpu)
3074 struct rq *rq = cpu_rq(cpu);
3075 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3076 unsigned long load_avg = rq->cfs.runnable_load_avg;
3079 return load_avg / nr_running;
3084 static void record_wakee(struct task_struct *p)
3087 * Rough decay (wiping) for cost saving, don't worry
3088 * about the boundary, really active task won't care
3091 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3092 current->wakee_flips = 0;
3093 current->wakee_flip_decay_ts = jiffies;
3096 if (current->last_wakee != p) {
3097 current->last_wakee = p;
3098 current->wakee_flips++;
3102 static void task_waking_fair(struct task_struct *p)
3104 struct sched_entity *se = &p->se;
3105 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3108 #ifndef CONFIG_64BIT
3109 u64 min_vruntime_copy;
3112 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3114 min_vruntime = cfs_rq->min_vruntime;
3115 } while (min_vruntime != min_vruntime_copy);
3117 min_vruntime = cfs_rq->min_vruntime;
3120 se->vruntime -= min_vruntime;
3124 #ifdef CONFIG_FAIR_GROUP_SCHED
3126 * effective_load() calculates the load change as seen from the root_task_group
3128 * Adding load to a group doesn't make a group heavier, but can cause movement
3129 * of group shares between cpus. Assuming the shares were perfectly aligned one
3130 * can calculate the shift in shares.
3132 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3133 * on this @cpu and results in a total addition (subtraction) of @wg to the
3134 * total group weight.
3136 * Given a runqueue weight distribution (rw_i) we can compute a shares
3137 * distribution (s_i) using:
3139 * s_i = rw_i / \Sum rw_j (1)
3141 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3142 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3143 * shares distribution (s_i):
3145 * rw_i = { 2, 4, 1, 0 }
3146 * s_i = { 2/7, 4/7, 1/7, 0 }
3148 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3149 * task used to run on and the CPU the waker is running on), we need to
3150 * compute the effect of waking a task on either CPU and, in case of a sync
3151 * wakeup, compute the effect of the current task going to sleep.
3153 * So for a change of @wl to the local @cpu with an overall group weight change
3154 * of @wl we can compute the new shares distribution (s'_i) using:
3156 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3158 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3159 * differences in waking a task to CPU 0. The additional task changes the
3160 * weight and shares distributions like:
3162 * rw'_i = { 3, 4, 1, 0 }
3163 * s'_i = { 3/8, 4/8, 1/8, 0 }
3165 * We can then compute the difference in effective weight by using:
3167 * dw_i = S * (s'_i - s_i) (3)
3169 * Where 'S' is the group weight as seen by its parent.
3171 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3172 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3173 * 4/7) times the weight of the group.
3175 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3177 struct sched_entity *se = tg->se[cpu];
3179 if (!tg->parent) /* the trivial, non-cgroup case */
3182 for_each_sched_entity(se) {
3188 * W = @wg + \Sum rw_j
3190 W = wg + calc_tg_weight(tg, se->my_q);
3195 w = se->my_q->load.weight + wl;
3198 * wl = S * s'_i; see (2)
3201 wl = (w * tg->shares) / W;
3206 * Per the above, wl is the new se->load.weight value; since
3207 * those are clipped to [MIN_SHARES, ...) do so now. See
3208 * calc_cfs_shares().
3210 if (wl < MIN_SHARES)
3214 * wl = dw_i = S * (s'_i - s_i); see (3)
3216 wl -= se->load.weight;
3219 * Recursively apply this logic to all parent groups to compute
3220 * the final effective load change on the root group. Since
3221 * only the @tg group gets extra weight, all parent groups can
3222 * only redistribute existing shares. @wl is the shift in shares
3223 * resulting from this level per the above.
3232 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3233 unsigned long wl, unsigned long wg)
3240 static int wake_wide(struct task_struct *p)
3242 int factor = this_cpu_read(sd_llc_size);
3245 * Yeah, it's the switching-frequency, could means many wakee or
3246 * rapidly switch, use factor here will just help to automatically
3247 * adjust the loose-degree, so bigger node will lead to more pull.
3249 if (p->wakee_flips > factor) {
3251 * wakee is somewhat hot, it needs certain amount of cpu
3252 * resource, so if waker is far more hot, prefer to leave
3255 if (current->wakee_flips > (factor * p->wakee_flips))
3262 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3264 s64 this_load, load;
3265 int idx, this_cpu, prev_cpu;
3266 unsigned long tl_per_task;
3267 struct task_group *tg;
3268 unsigned long weight;
3272 * If we wake multiple tasks be careful to not bounce
3273 * ourselves around too much.
3279 this_cpu = smp_processor_id();
3280 prev_cpu = task_cpu(p);
3281 load = source_load(prev_cpu, idx);
3282 this_load = target_load(this_cpu, idx);
3285 * If sync wakeup then subtract the (maximum possible)
3286 * effect of the currently running task from the load
3287 * of the current CPU:
3290 tg = task_group(current);
3291 weight = current->se.load.weight;
3293 this_load += effective_load(tg, this_cpu, -weight, -weight);
3294 load += effective_load(tg, prev_cpu, 0, -weight);
3298 weight = p->se.load.weight;
3301 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3302 * due to the sync cause above having dropped this_load to 0, we'll
3303 * always have an imbalance, but there's really nothing you can do
3304 * about that, so that's good too.
3306 * Otherwise check if either cpus are near enough in load to allow this
3307 * task to be woken on this_cpu.
3309 if (this_load > 0) {
3310 s64 this_eff_load, prev_eff_load;
3312 this_eff_load = 100;
3313 this_eff_load *= power_of(prev_cpu);
3314 this_eff_load *= this_load +
3315 effective_load(tg, this_cpu, weight, weight);
3317 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3318 prev_eff_load *= power_of(this_cpu);
3319 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3321 balanced = this_eff_load <= prev_eff_load;
3326 * If the currently running task will sleep within
3327 * a reasonable amount of time then attract this newly
3330 if (sync && balanced)
3333 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3334 tl_per_task = cpu_avg_load_per_task(this_cpu);
3337 (this_load <= load &&
3338 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3340 * This domain has SD_WAKE_AFFINE and
3341 * p is cache cold in this domain, and
3342 * there is no bad imbalance.
3344 schedstat_inc(sd, ttwu_move_affine);
3345 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3353 * find_idlest_group finds and returns the least busy CPU group within the
3356 static struct sched_group *
3357 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3358 int this_cpu, int load_idx)
3360 struct sched_group *idlest = NULL, *group = sd->groups;
3361 unsigned long min_load = ULONG_MAX, this_load = 0;
3362 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3365 unsigned long load, avg_load;
3369 /* Skip over this group if it has no CPUs allowed */
3370 if (!cpumask_intersects(sched_group_cpus(group),
3371 tsk_cpus_allowed(p)))
3374 local_group = cpumask_test_cpu(this_cpu,
3375 sched_group_cpus(group));
3377 /* Tally up the load of all CPUs in the group */
3380 for_each_cpu(i, sched_group_cpus(group)) {
3381 /* Bias balancing toward cpus of our domain */
3383 load = source_load(i, load_idx);
3385 load = target_load(i, load_idx);
3390 /* Adjust by relative CPU power of the group */
3391 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3394 this_load = avg_load;
3395 } else if (avg_load < min_load) {
3396 min_load = avg_load;
3399 } while (group = group->next, group != sd->groups);
3401 if (!idlest || 100*this_load < imbalance*min_load)
3407 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3410 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3412 unsigned long load, min_load = ULONG_MAX;
3416 /* Traverse only the allowed CPUs */
3417 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3418 load = weighted_cpuload(i);
3420 if (load < min_load || (load == min_load && i == this_cpu)) {
3430 * Try and locate an idle CPU in the sched_domain.
3432 static int select_idle_sibling(struct task_struct *p, int target)
3434 struct sched_domain *sd;
3435 struct sched_group *sg;
3436 int i = task_cpu(p);
3438 if (idle_cpu(target))
3442 * If the prevous cpu is cache affine and idle, don't be stupid.
3444 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3448 * Otherwise, iterate the domains and find an elegible idle cpu.
3450 sd = rcu_dereference(per_cpu(sd_llc, target));
3451 for_each_lower_domain(sd) {
3454 if (!cpumask_intersects(sched_group_cpus(sg),
3455 tsk_cpus_allowed(p)))
3458 for_each_cpu(i, sched_group_cpus(sg)) {
3459 if (i == target || !idle_cpu(i))
3463 target = cpumask_first_and(sched_group_cpus(sg),
3464 tsk_cpus_allowed(p));
3468 } while (sg != sd->groups);
3475 * sched_balance_self: balance the current task (running on cpu) in domains
3476 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3479 * Balance, ie. select the least loaded group.
3481 * Returns the target CPU number, or the same CPU if no balancing is needed.
3483 * preempt must be disabled.
3486 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3488 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3489 int cpu = smp_processor_id();
3490 int prev_cpu = task_cpu(p);
3492 int want_affine = 0;
3493 int sync = wake_flags & WF_SYNC;
3495 if (p->nr_cpus_allowed == 1)
3498 if (sd_flag & SD_BALANCE_WAKE) {
3499 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3505 for_each_domain(cpu, tmp) {
3506 if (!(tmp->flags & SD_LOAD_BALANCE))
3510 * If both cpu and prev_cpu are part of this domain,
3511 * cpu is a valid SD_WAKE_AFFINE target.
3513 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3514 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3519 if (tmp->flags & sd_flag)
3524 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3527 new_cpu = select_idle_sibling(p, prev_cpu);
3532 int load_idx = sd->forkexec_idx;
3533 struct sched_group *group;
3536 if (!(sd->flags & sd_flag)) {
3541 if (sd_flag & SD_BALANCE_WAKE)
3542 load_idx = sd->wake_idx;
3544 group = find_idlest_group(sd, p, cpu, load_idx);
3550 new_cpu = find_idlest_cpu(group, p, cpu);
3551 if (new_cpu == -1 || new_cpu == cpu) {
3552 /* Now try balancing at a lower domain level of cpu */
3557 /* Now try balancing at a lower domain level of new_cpu */
3559 weight = sd->span_weight;
3561 for_each_domain(cpu, tmp) {
3562 if (weight <= tmp->span_weight)
3564 if (tmp->flags & sd_flag)
3567 /* while loop will break here if sd == NULL */
3576 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3577 * cfs_rq_of(p) references at time of call are still valid and identify the
3578 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3579 * other assumptions, including the state of rq->lock, should be made.
3582 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3584 struct sched_entity *se = &p->se;
3585 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3588 * Load tracking: accumulate removed load so that it can be processed
3589 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3590 * to blocked load iff they have a positive decay-count. It can never
3591 * be negative here since on-rq tasks have decay-count == 0.
3593 if (se->avg.decay_count) {
3594 se->avg.decay_count = -__synchronize_entity_decay(se);
3595 atomic_long_add(se->avg.load_avg_contrib,
3596 &cfs_rq->removed_load);
3599 #endif /* CONFIG_SMP */
3601 static unsigned long
3602 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3604 unsigned long gran = sysctl_sched_wakeup_granularity;
3607 * Since its curr running now, convert the gran from real-time
3608 * to virtual-time in his units.
3610 * By using 'se' instead of 'curr' we penalize light tasks, so
3611 * they get preempted easier. That is, if 'se' < 'curr' then
3612 * the resulting gran will be larger, therefore penalizing the
3613 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3614 * be smaller, again penalizing the lighter task.
3616 * This is especially important for buddies when the leftmost
3617 * task is higher priority than the buddy.
3619 return calc_delta_fair(gran, se);
3623 * Should 'se' preempt 'curr'.
3637 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3639 s64 gran, vdiff = curr->vruntime - se->vruntime;
3644 gran = wakeup_gran(curr, se);
3651 static void set_last_buddy(struct sched_entity *se)
3653 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3656 for_each_sched_entity(se)
3657 cfs_rq_of(se)->last = se;
3660 static void set_next_buddy(struct sched_entity *se)
3662 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3665 for_each_sched_entity(se)
3666 cfs_rq_of(se)->next = se;
3669 static void set_skip_buddy(struct sched_entity *se)
3671 for_each_sched_entity(se)
3672 cfs_rq_of(se)->skip = se;
3676 * Preempt the current task with a newly woken task if needed:
3678 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3680 struct task_struct *curr = rq->curr;
3681 struct sched_entity *se = &curr->se, *pse = &p->se;
3682 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3683 int scale = cfs_rq->nr_running >= sched_nr_latency;
3684 int next_buddy_marked = 0;
3686 if (unlikely(se == pse))
3690 * This is possible from callers such as move_task(), in which we
3691 * unconditionally check_prempt_curr() after an enqueue (which may have
3692 * lead to a throttle). This both saves work and prevents false
3693 * next-buddy nomination below.
3695 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3698 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3699 set_next_buddy(pse);
3700 next_buddy_marked = 1;
3704 * We can come here with TIF_NEED_RESCHED already set from new task
3707 * Note: this also catches the edge-case of curr being in a throttled
3708 * group (e.g. via set_curr_task), since update_curr() (in the
3709 * enqueue of curr) will have resulted in resched being set. This
3710 * prevents us from potentially nominating it as a false LAST_BUDDY
3713 if (test_tsk_need_resched(curr))
3716 /* Idle tasks are by definition preempted by non-idle tasks. */
3717 if (unlikely(curr->policy == SCHED_IDLE) &&
3718 likely(p->policy != SCHED_IDLE))
3722 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3723 * is driven by the tick):
3725 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3728 find_matching_se(&se, &pse);
3729 update_curr(cfs_rq_of(se));
3731 if (wakeup_preempt_entity(se, pse) == 1) {
3733 * Bias pick_next to pick the sched entity that is
3734 * triggering this preemption.
3736 if (!next_buddy_marked)
3737 set_next_buddy(pse);
3746 * Only set the backward buddy when the current task is still
3747 * on the rq. This can happen when a wakeup gets interleaved
3748 * with schedule on the ->pre_schedule() or idle_balance()
3749 * point, either of which can * drop the rq lock.
3751 * Also, during early boot the idle thread is in the fair class,
3752 * for obvious reasons its a bad idea to schedule back to it.
3754 if (unlikely(!se->on_rq || curr == rq->idle))
3757 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3761 static struct task_struct *pick_next_task_fair(struct rq *rq)
3763 struct task_struct *p;
3764 struct cfs_rq *cfs_rq = &rq->cfs;
3765 struct sched_entity *se;
3767 if (!cfs_rq->nr_running)
3771 se = pick_next_entity(cfs_rq);
3772 set_next_entity(cfs_rq, se);
3773 cfs_rq = group_cfs_rq(se);
3777 if (hrtick_enabled(rq))
3778 hrtick_start_fair(rq, p);
3784 * Account for a descheduled task:
3786 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3788 struct sched_entity *se = &prev->se;
3789 struct cfs_rq *cfs_rq;
3791 for_each_sched_entity(se) {
3792 cfs_rq = cfs_rq_of(se);
3793 put_prev_entity(cfs_rq, se);
3798 * sched_yield() is very simple
3800 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3802 static void yield_task_fair(struct rq *rq)
3804 struct task_struct *curr = rq->curr;
3805 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3806 struct sched_entity *se = &curr->se;
3809 * Are we the only task in the tree?
3811 if (unlikely(rq->nr_running == 1))
3814 clear_buddies(cfs_rq, se);
3816 if (curr->policy != SCHED_BATCH) {
3817 update_rq_clock(rq);
3819 * Update run-time statistics of the 'current'.
3821 update_curr(cfs_rq);
3823 * Tell update_rq_clock() that we've just updated,
3824 * so we don't do microscopic update in schedule()
3825 * and double the fastpath cost.
3827 rq->skip_clock_update = 1;
3833 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3835 struct sched_entity *se = &p->se;
3837 /* throttled hierarchies are not runnable */
3838 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3841 /* Tell the scheduler that we'd really like pse to run next. */
3844 yield_task_fair(rq);
3850 /**************************************************
3851 * Fair scheduling class load-balancing methods.
3855 * The purpose of load-balancing is to achieve the same basic fairness the
3856 * per-cpu scheduler provides, namely provide a proportional amount of compute
3857 * time to each task. This is expressed in the following equation:
3859 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3861 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3862 * W_i,0 is defined as:
3864 * W_i,0 = \Sum_j w_i,j (2)
3866 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3867 * is derived from the nice value as per prio_to_weight[].
3869 * The weight average is an exponential decay average of the instantaneous
3872 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3874 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3875 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3876 * can also include other factors [XXX].
3878 * To achieve this balance we define a measure of imbalance which follows
3879 * directly from (1):
3881 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3883 * We them move tasks around to minimize the imbalance. In the continuous
3884 * function space it is obvious this converges, in the discrete case we get
3885 * a few fun cases generally called infeasible weight scenarios.
3888 * - infeasible weights;
3889 * - local vs global optima in the discrete case. ]
3894 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3895 * for all i,j solution, we create a tree of cpus that follows the hardware
3896 * topology where each level pairs two lower groups (or better). This results
3897 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3898 * tree to only the first of the previous level and we decrease the frequency
3899 * of load-balance at each level inv. proportional to the number of cpus in
3905 * \Sum { --- * --- * 2^i } = O(n) (5)
3907 * `- size of each group
3908 * | | `- number of cpus doing load-balance
3910 * `- sum over all levels
3912 * Coupled with a limit on how many tasks we can migrate every balance pass,
3913 * this makes (5) the runtime complexity of the balancer.
3915 * An important property here is that each CPU is still (indirectly) connected
3916 * to every other cpu in at most O(log n) steps:
3918 * The adjacency matrix of the resulting graph is given by:
3921 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3924 * And you'll find that:
3926 * A^(log_2 n)_i,j != 0 for all i,j (7)
3928 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3929 * The task movement gives a factor of O(m), giving a convergence complexity
3932 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3937 * In order to avoid CPUs going idle while there's still work to do, new idle
3938 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3939 * tree itself instead of relying on other CPUs to bring it work.
3941 * This adds some complexity to both (5) and (8) but it reduces the total idle
3949 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3952 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3957 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3959 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3961 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3964 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3965 * rewrite all of this once again.]
3968 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3970 #define LBF_ALL_PINNED 0x01
3971 #define LBF_NEED_BREAK 0x02
3972 #define LBF_DST_PINNED 0x04
3973 #define LBF_SOME_PINNED 0x08
3976 struct sched_domain *sd;
3984 struct cpumask *dst_grpmask;
3986 enum cpu_idle_type idle;
3988 /* The set of CPUs under consideration for load-balancing */
3989 struct cpumask *cpus;
3994 unsigned int loop_break;
3995 unsigned int loop_max;
3999 * move_task - move a task from one runqueue to another runqueue.
4000 * Both runqueues must be locked.
4002 static void move_task(struct task_struct *p, struct lb_env *env)
4004 deactivate_task(env->src_rq, p, 0);
4005 set_task_cpu(p, env->dst_cpu);
4006 activate_task(env->dst_rq, p, 0);
4007 check_preempt_curr(env->dst_rq, p, 0);
4011 * Is this task likely cache-hot:
4014 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4018 if (p->sched_class != &fair_sched_class)
4021 if (unlikely(p->policy == SCHED_IDLE))
4025 * Buddy candidates are cache hot:
4027 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4028 (&p->se == cfs_rq_of(&p->se)->next ||
4029 &p->se == cfs_rq_of(&p->se)->last))
4032 if (sysctl_sched_migration_cost == -1)
4034 if (sysctl_sched_migration_cost == 0)
4037 delta = now - p->se.exec_start;
4039 return delta < (s64)sysctl_sched_migration_cost;
4043 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4046 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4048 int tsk_cache_hot = 0;
4050 * We do not migrate tasks that are:
4051 * 1) throttled_lb_pair, or
4052 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4053 * 3) running (obviously), or
4054 * 4) are cache-hot on their current CPU.
4056 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4059 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4062 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4064 env->flags |= LBF_SOME_PINNED;
4067 * Remember if this task can be migrated to any other cpu in
4068 * our sched_group. We may want to revisit it if we couldn't
4069 * meet load balance goals by pulling other tasks on src_cpu.
4071 * Also avoid computing new_dst_cpu if we have already computed
4072 * one in current iteration.
4074 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4077 /* Prevent to re-select dst_cpu via env's cpus */
4078 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4079 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4080 env->flags |= LBF_DST_PINNED;
4081 env->new_dst_cpu = cpu;
4089 /* Record that we found atleast one task that could run on dst_cpu */
4090 env->flags &= ~LBF_ALL_PINNED;
4092 if (task_running(env->src_rq, p)) {
4093 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4098 * Aggressive migration if:
4099 * 1) task is cache cold, or
4100 * 2) too many balance attempts have failed.
4103 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4104 if (!tsk_cache_hot ||
4105 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4107 if (tsk_cache_hot) {
4108 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4109 schedstat_inc(p, se.statistics.nr_forced_migrations);
4115 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4120 * move_one_task tries to move exactly one task from busiest to this_rq, as
4121 * part of active balancing operations within "domain".
4122 * Returns 1 if successful and 0 otherwise.
4124 * Called with both runqueues locked.
4126 static int move_one_task(struct lb_env *env)
4128 struct task_struct *p, *n;
4130 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4131 if (!can_migrate_task(p, env))
4136 * Right now, this is only the second place move_task()
4137 * is called, so we can safely collect move_task()
4138 * stats here rather than inside move_task().
4140 schedstat_inc(env->sd, lb_gained[env->idle]);
4146 static unsigned long task_h_load(struct task_struct *p);
4148 static const unsigned int sched_nr_migrate_break = 32;
4151 * move_tasks tries to move up to imbalance weighted load from busiest to
4152 * this_rq, as part of a balancing operation within domain "sd".
4153 * Returns 1 if successful and 0 otherwise.
4155 * Called with both runqueues locked.
4157 static int move_tasks(struct lb_env *env)
4159 struct list_head *tasks = &env->src_rq->cfs_tasks;
4160 struct task_struct *p;
4164 if (env->imbalance <= 0)
4167 while (!list_empty(tasks)) {
4168 p = list_first_entry(tasks, struct task_struct, se.group_node);
4171 /* We've more or less seen every task there is, call it quits */
4172 if (env->loop > env->loop_max)
4175 /* take a breather every nr_migrate tasks */
4176 if (env->loop > env->loop_break) {
4177 env->loop_break += sched_nr_migrate_break;
4178 env->flags |= LBF_NEED_BREAK;
4182 if (!can_migrate_task(p, env))
4185 load = task_h_load(p);
4187 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4190 if ((load / 2) > env->imbalance)
4195 env->imbalance -= load;
4197 #ifdef CONFIG_PREEMPT
4199 * NEWIDLE balancing is a source of latency, so preemptible
4200 * kernels will stop after the first task is pulled to minimize
4201 * the critical section.
4203 if (env->idle == CPU_NEWLY_IDLE)
4208 * We only want to steal up to the prescribed amount of
4211 if (env->imbalance <= 0)
4216 list_move_tail(&p->se.group_node, tasks);
4220 * Right now, this is one of only two places move_task() is called,
4221 * so we can safely collect move_task() stats here rather than
4222 * inside move_task().
4224 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4229 #ifdef CONFIG_FAIR_GROUP_SCHED
4231 * update tg->load_weight by folding this cpu's load_avg
4233 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4235 struct sched_entity *se = tg->se[cpu];
4236 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4238 /* throttled entities do not contribute to load */
4239 if (throttled_hierarchy(cfs_rq))
4242 update_cfs_rq_blocked_load(cfs_rq, 1);
4245 update_entity_load_avg(se, 1);
4247 * We pivot on our runnable average having decayed to zero for
4248 * list removal. This generally implies that all our children
4249 * have also been removed (modulo rounding error or bandwidth
4250 * control); however, such cases are rare and we can fix these
4253 * TODO: fix up out-of-order children on enqueue.
4255 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4256 list_del_leaf_cfs_rq(cfs_rq);
4258 struct rq *rq = rq_of(cfs_rq);
4259 update_rq_runnable_avg(rq, rq->nr_running);
4263 static void update_blocked_averages(int cpu)
4265 struct rq *rq = cpu_rq(cpu);
4266 struct cfs_rq *cfs_rq;
4267 unsigned long flags;
4269 raw_spin_lock_irqsave(&rq->lock, flags);
4270 update_rq_clock(rq);
4272 * Iterates the task_group tree in a bottom up fashion, see
4273 * list_add_leaf_cfs_rq() for details.
4275 for_each_leaf_cfs_rq(rq, cfs_rq) {
4277 * Note: We may want to consider periodically releasing
4278 * rq->lock about these updates so that creating many task
4279 * groups does not result in continually extending hold time.
4281 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4284 raw_spin_unlock_irqrestore(&rq->lock, flags);
4288 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4289 * This needs to be done in a top-down fashion because the load of a child
4290 * group is a fraction of its parents load.
4292 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4294 struct rq *rq = rq_of(cfs_rq);
4295 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4296 unsigned long now = jiffies;
4299 if (cfs_rq->last_h_load_update == now)
4302 cfs_rq->h_load_next = NULL;
4303 for_each_sched_entity(se) {
4304 cfs_rq = cfs_rq_of(se);
4305 cfs_rq->h_load_next = se;
4306 if (cfs_rq->last_h_load_update == now)
4311 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4312 cfs_rq->last_h_load_update = now;
4315 while ((se = cfs_rq->h_load_next) != NULL) {
4316 load = cfs_rq->h_load;
4317 load = div64_ul(load * se->avg.load_avg_contrib,
4318 cfs_rq->runnable_load_avg + 1);
4319 cfs_rq = group_cfs_rq(se);
4320 cfs_rq->h_load = load;
4321 cfs_rq->last_h_load_update = now;
4325 static unsigned long task_h_load(struct task_struct *p)
4327 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4329 update_cfs_rq_h_load(cfs_rq);
4330 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4331 cfs_rq->runnable_load_avg + 1);
4334 static inline void update_blocked_averages(int cpu)
4338 static unsigned long task_h_load(struct task_struct *p)
4340 return p->se.avg.load_avg_contrib;
4344 /********** Helpers for find_busiest_group ************************/
4346 * sg_lb_stats - stats of a sched_group required for load_balancing
4348 struct sg_lb_stats {
4349 unsigned long avg_load; /*Avg load across the CPUs of the group */
4350 unsigned long group_load; /* Total load over the CPUs of the group */
4351 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4352 unsigned long load_per_task;
4353 unsigned long group_power;
4354 unsigned int sum_nr_running; /* Nr tasks running in the group */
4355 unsigned int group_capacity;
4356 unsigned int idle_cpus;
4357 unsigned int group_weight;
4358 int group_imb; /* Is there an imbalance in the group ? */
4359 int group_has_capacity; /* Is there extra capacity in the group? */
4363 * sd_lb_stats - Structure to store the statistics of a sched_domain
4364 * during load balancing.
4366 struct sd_lb_stats {
4367 struct sched_group *busiest; /* Busiest group in this sd */
4368 struct sched_group *local; /* Local group in this sd */
4369 unsigned long total_load; /* Total load of all groups in sd */
4370 unsigned long total_pwr; /* Total power of all groups in sd */
4371 unsigned long avg_load; /* Average load across all groups in sd */
4373 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4374 struct sg_lb_stats local_stat; /* Statistics of the local group */
4377 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4380 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4381 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4382 * We must however clear busiest_stat::avg_load because
4383 * update_sd_pick_busiest() reads this before assignment.
4385 *sds = (struct sd_lb_stats){
4397 * get_sd_load_idx - Obtain the load index for a given sched domain.
4398 * @sd: The sched_domain whose load_idx is to be obtained.
4399 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4401 * Return: The load index.
4403 static inline int get_sd_load_idx(struct sched_domain *sd,
4404 enum cpu_idle_type idle)
4410 load_idx = sd->busy_idx;
4413 case CPU_NEWLY_IDLE:
4414 load_idx = sd->newidle_idx;
4417 load_idx = sd->idle_idx;
4424 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4426 return SCHED_POWER_SCALE;
4429 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4431 return default_scale_freq_power(sd, cpu);
4434 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4436 unsigned long weight = sd->span_weight;
4437 unsigned long smt_gain = sd->smt_gain;
4444 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4446 return default_scale_smt_power(sd, cpu);
4449 static unsigned long scale_rt_power(int cpu)
4451 struct rq *rq = cpu_rq(cpu);
4452 u64 total, available, age_stamp, avg;
4455 * Since we're reading these variables without serialization make sure
4456 * we read them once before doing sanity checks on them.
4458 age_stamp = ACCESS_ONCE(rq->age_stamp);
4459 avg = ACCESS_ONCE(rq->rt_avg);
4461 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4463 if (unlikely(total < avg)) {
4464 /* Ensures that power won't end up being negative */
4467 available = total - avg;
4470 if (unlikely((s64)total < SCHED_POWER_SCALE))
4471 total = SCHED_POWER_SCALE;
4473 total >>= SCHED_POWER_SHIFT;
4475 return div_u64(available, total);
4478 static void update_cpu_power(struct sched_domain *sd, int cpu)
4480 unsigned long weight = sd->span_weight;
4481 unsigned long power = SCHED_POWER_SCALE;
4482 struct sched_group *sdg = sd->groups;
4484 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4485 if (sched_feat(ARCH_POWER))
4486 power *= arch_scale_smt_power(sd, cpu);
4488 power *= default_scale_smt_power(sd, cpu);
4490 power >>= SCHED_POWER_SHIFT;
4493 sdg->sgp->power_orig = power;
4495 if (sched_feat(ARCH_POWER))
4496 power *= arch_scale_freq_power(sd, cpu);
4498 power *= default_scale_freq_power(sd, cpu);
4500 power >>= SCHED_POWER_SHIFT;
4502 power *= scale_rt_power(cpu);
4503 power >>= SCHED_POWER_SHIFT;
4508 cpu_rq(cpu)->cpu_power = power;
4509 sdg->sgp->power = power;
4512 void update_group_power(struct sched_domain *sd, int cpu)
4514 struct sched_domain *child = sd->child;
4515 struct sched_group *group, *sdg = sd->groups;
4516 unsigned long power, power_orig;
4517 unsigned long interval;
4519 interval = msecs_to_jiffies(sd->balance_interval);
4520 interval = clamp(interval, 1UL, max_load_balance_interval);
4521 sdg->sgp->next_update = jiffies + interval;
4524 update_cpu_power(sd, cpu);
4528 power_orig = power = 0;
4530 if (child->flags & SD_OVERLAP) {
4532 * SD_OVERLAP domains cannot assume that child groups
4533 * span the current group.
4536 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4537 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4539 power_orig += sg->sgp->power_orig;
4540 power += sg->sgp->power;
4544 * !SD_OVERLAP domains can assume that child groups
4545 * span the current group.
4548 group = child->groups;
4550 power_orig += group->sgp->power_orig;
4551 power += group->sgp->power;
4552 group = group->next;
4553 } while (group != child->groups);
4556 sdg->sgp->power_orig = power_orig;
4557 sdg->sgp->power = power;
4561 * Try and fix up capacity for tiny siblings, this is needed when
4562 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4563 * which on its own isn't powerful enough.
4565 * See update_sd_pick_busiest() and check_asym_packing().
4568 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4571 * Only siblings can have significantly less than SCHED_POWER_SCALE
4573 if (!(sd->flags & SD_SHARE_CPUPOWER))
4577 * If ~90% of the cpu_power is still there, we're good.
4579 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4586 * Group imbalance indicates (and tries to solve) the problem where balancing
4587 * groups is inadequate due to tsk_cpus_allowed() constraints.
4589 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4590 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4593 * { 0 1 2 3 } { 4 5 6 7 }
4596 * If we were to balance group-wise we'd place two tasks in the first group and
4597 * two tasks in the second group. Clearly this is undesired as it will overload
4598 * cpu 3 and leave one of the cpus in the second group unused.
4600 * The current solution to this issue is detecting the skew in the first group
4601 * by noticing the lower domain failed to reach balance and had difficulty
4602 * moving tasks due to affinity constraints.
4604 * When this is so detected; this group becomes a candidate for busiest; see
4605 * update_sd_pick_busiest(). And calculcate_imbalance() and
4606 * find_busiest_group() avoid some of the usual balance conditions to allow it
4607 * to create an effective group imbalance.
4609 * This is a somewhat tricky proposition since the next run might not find the
4610 * group imbalance and decide the groups need to be balanced again. A most
4611 * subtle and fragile situation.
4614 static inline int sg_imbalanced(struct sched_group *group)
4616 return group->sgp->imbalance;
4620 * Compute the group capacity.
4622 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4623 * first dividing out the smt factor and computing the actual number of cores
4624 * and limit power unit capacity with that.
4626 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4628 unsigned int capacity, smt, cpus;
4629 unsigned int power, power_orig;
4631 power = group->sgp->power;
4632 power_orig = group->sgp->power_orig;
4633 cpus = group->group_weight;
4635 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4636 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4637 capacity = cpus / smt; /* cores */
4639 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4641 capacity = fix_small_capacity(env->sd, group);
4647 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4648 * @env: The load balancing environment.
4649 * @group: sched_group whose statistics are to be updated.
4650 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4651 * @local_group: Does group contain this_cpu.
4652 * @sgs: variable to hold the statistics for this group.
4654 static inline void update_sg_lb_stats(struct lb_env *env,
4655 struct sched_group *group, int load_idx,
4656 int local_group, struct sg_lb_stats *sgs)
4658 unsigned long nr_running;
4662 memset(sgs, 0, sizeof(*sgs));
4664 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4665 struct rq *rq = cpu_rq(i);
4667 nr_running = rq->nr_running;
4669 /* Bias balancing toward cpus of our domain */
4671 load = target_load(i, load_idx);
4673 load = source_load(i, load_idx);
4675 sgs->group_load += load;
4676 sgs->sum_nr_running += nr_running;
4677 sgs->sum_weighted_load += weighted_cpuload(i);
4682 /* Adjust by relative CPU power of the group */
4683 sgs->group_power = group->sgp->power;
4684 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4686 if (sgs->sum_nr_running)
4687 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4689 sgs->group_weight = group->group_weight;
4691 sgs->group_imb = sg_imbalanced(group);
4692 sgs->group_capacity = sg_capacity(env, group);
4694 if (sgs->group_capacity > sgs->sum_nr_running)
4695 sgs->group_has_capacity = 1;
4699 * update_sd_pick_busiest - return 1 on busiest group
4700 * @env: The load balancing environment.
4701 * @sds: sched_domain statistics
4702 * @sg: sched_group candidate to be checked for being the busiest
4703 * @sgs: sched_group statistics
4705 * Determine if @sg is a busier group than the previously selected
4708 * Return: %true if @sg is a busier group than the previously selected
4709 * busiest group. %false otherwise.
4711 static bool update_sd_pick_busiest(struct lb_env *env,
4712 struct sd_lb_stats *sds,
4713 struct sched_group *sg,
4714 struct sg_lb_stats *sgs)
4716 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4719 if (sgs->sum_nr_running > sgs->group_capacity)
4726 * ASYM_PACKING needs to move all the work to the lowest
4727 * numbered CPUs in the group, therefore mark all groups
4728 * higher than ourself as busy.
4730 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4731 env->dst_cpu < group_first_cpu(sg)) {
4735 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4743 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4744 * @env: The load balancing environment.
4745 * @balance: Should we balance.
4746 * @sds: variable to hold the statistics for this sched_domain.
4748 static inline void update_sd_lb_stats(struct lb_env *env,
4749 struct sd_lb_stats *sds)
4751 struct sched_domain *child = env->sd->child;
4752 struct sched_group *sg = env->sd->groups;
4753 struct sg_lb_stats tmp_sgs;
4754 int load_idx, prefer_sibling = 0;
4756 if (child && child->flags & SD_PREFER_SIBLING)
4759 load_idx = get_sd_load_idx(env->sd, env->idle);
4762 struct sg_lb_stats *sgs = &tmp_sgs;
4765 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4768 sgs = &sds->local_stat;
4770 if (env->idle != CPU_NEWLY_IDLE ||
4771 time_after_eq(jiffies, sg->sgp->next_update))
4772 update_group_power(env->sd, env->dst_cpu);
4775 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4781 * In case the child domain prefers tasks go to siblings
4782 * first, lower the sg capacity to one so that we'll try
4783 * and move all the excess tasks away. We lower the capacity
4784 * of a group only if the local group has the capacity to fit
4785 * these excess tasks, i.e. nr_running < group_capacity. The
4786 * extra check prevents the case where you always pull from the
4787 * heaviest group when it is already under-utilized (possible
4788 * with a large weight task outweighs the tasks on the system).
4790 if (prefer_sibling && sds->local &&
4791 sds->local_stat.group_has_capacity)
4792 sgs->group_capacity = min(sgs->group_capacity, 1U);
4794 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4796 sds->busiest_stat = *sgs;
4800 /* Now, start updating sd_lb_stats */
4801 sds->total_load += sgs->group_load;
4802 sds->total_pwr += sgs->group_power;
4805 } while (sg != env->sd->groups);
4809 * check_asym_packing - Check to see if the group is packed into the
4812 * This is primarily intended to used at the sibling level. Some
4813 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4814 * case of POWER7, it can move to lower SMT modes only when higher
4815 * threads are idle. When in lower SMT modes, the threads will
4816 * perform better since they share less core resources. Hence when we
4817 * have idle threads, we want them to be the higher ones.
4819 * This packing function is run on idle threads. It checks to see if
4820 * the busiest CPU in this domain (core in the P7 case) has a higher
4821 * CPU number than the packing function is being run on. Here we are
4822 * assuming lower CPU number will be equivalent to lower a SMT thread
4825 * Return: 1 when packing is required and a task should be moved to
4826 * this CPU. The amount of the imbalance is returned in *imbalance.
4828 * @env: The load balancing environment.
4829 * @sds: Statistics of the sched_domain which is to be packed
4831 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4835 if (!(env->sd->flags & SD_ASYM_PACKING))
4841 busiest_cpu = group_first_cpu(sds->busiest);
4842 if (env->dst_cpu > busiest_cpu)
4845 env->imbalance = DIV_ROUND_CLOSEST(
4846 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4853 * fix_small_imbalance - Calculate the minor imbalance that exists
4854 * amongst the groups of a sched_domain, during
4856 * @env: The load balancing environment.
4857 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4860 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4862 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4863 unsigned int imbn = 2;
4864 unsigned long scaled_busy_load_per_task;
4865 struct sg_lb_stats *local, *busiest;
4867 local = &sds->local_stat;
4868 busiest = &sds->busiest_stat;
4870 if (!local->sum_nr_running)
4871 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4872 else if (busiest->load_per_task > local->load_per_task)
4875 scaled_busy_load_per_task =
4876 (busiest->load_per_task * SCHED_POWER_SCALE) /
4877 busiest->group_power;
4879 if (busiest->avg_load + scaled_busy_load_per_task >=
4880 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4881 env->imbalance = busiest->load_per_task;
4886 * OK, we don't have enough imbalance to justify moving tasks,
4887 * however we may be able to increase total CPU power used by
4891 pwr_now += busiest->group_power *
4892 min(busiest->load_per_task, busiest->avg_load);
4893 pwr_now += local->group_power *
4894 min(local->load_per_task, local->avg_load);
4895 pwr_now /= SCHED_POWER_SCALE;
4897 /* Amount of load we'd subtract */
4898 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4899 busiest->group_power;
4900 if (busiest->avg_load > tmp) {
4901 pwr_move += busiest->group_power *
4902 min(busiest->load_per_task,
4903 busiest->avg_load - tmp);
4906 /* Amount of load we'd add */
4907 if (busiest->avg_load * busiest->group_power <
4908 busiest->load_per_task * SCHED_POWER_SCALE) {
4909 tmp = (busiest->avg_load * busiest->group_power) /
4912 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4915 pwr_move += local->group_power *
4916 min(local->load_per_task, local->avg_load + tmp);
4917 pwr_move /= SCHED_POWER_SCALE;
4919 /* Move if we gain throughput */
4920 if (pwr_move > pwr_now)
4921 env->imbalance = busiest->load_per_task;
4925 * calculate_imbalance - Calculate the amount of imbalance present within the
4926 * groups of a given sched_domain during load balance.
4927 * @env: load balance environment
4928 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4930 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4932 unsigned long max_pull, load_above_capacity = ~0UL;
4933 struct sg_lb_stats *local, *busiest;
4935 local = &sds->local_stat;
4936 busiest = &sds->busiest_stat;
4938 if (busiest->group_imb) {
4940 * In the group_imb case we cannot rely on group-wide averages
4941 * to ensure cpu-load equilibrium, look at wider averages. XXX
4943 busiest->load_per_task =
4944 min(busiest->load_per_task, sds->avg_load);
4948 * In the presence of smp nice balancing, certain scenarios can have
4949 * max load less than avg load(as we skip the groups at or below
4950 * its cpu_power, while calculating max_load..)
4952 if (busiest->avg_load <= sds->avg_load ||
4953 local->avg_load >= sds->avg_load) {
4955 return fix_small_imbalance(env, sds);
4958 if (!busiest->group_imb) {
4960 * Don't want to pull so many tasks that a group would go idle.
4961 * Except of course for the group_imb case, since then we might
4962 * have to drop below capacity to reach cpu-load equilibrium.
4964 load_above_capacity =
4965 (busiest->sum_nr_running - busiest->group_capacity);
4967 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4968 load_above_capacity /= busiest->group_power;
4972 * We're trying to get all the cpus to the average_load, so we don't
4973 * want to push ourselves above the average load, nor do we wish to
4974 * reduce the max loaded cpu below the average load. At the same time,
4975 * we also don't want to reduce the group load below the group capacity
4976 * (so that we can implement power-savings policies etc). Thus we look
4977 * for the minimum possible imbalance.
4979 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4981 /* How much load to actually move to equalise the imbalance */
4982 env->imbalance = min(
4983 max_pull * busiest->group_power,
4984 (sds->avg_load - local->avg_load) * local->group_power
4985 ) / SCHED_POWER_SCALE;
4988 * if *imbalance is less than the average load per runnable task
4989 * there is no guarantee that any tasks will be moved so we'll have
4990 * a think about bumping its value to force at least one task to be
4993 if (env->imbalance < busiest->load_per_task)
4994 return fix_small_imbalance(env, sds);
4997 /******* find_busiest_group() helpers end here *********************/
5000 * find_busiest_group - Returns the busiest group within the sched_domain
5001 * if there is an imbalance. If there isn't an imbalance, and
5002 * the user has opted for power-savings, it returns a group whose
5003 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5004 * such a group exists.
5006 * Also calculates the amount of weighted load which should be moved
5007 * to restore balance.
5009 * @env: The load balancing environment.
5011 * Return: - The busiest group if imbalance exists.
5012 * - If no imbalance and user has opted for power-savings balance,
5013 * return the least loaded group whose CPUs can be
5014 * put to idle by rebalancing its tasks onto our group.
5016 static struct sched_group *find_busiest_group(struct lb_env *env)
5018 struct sg_lb_stats *local, *busiest;
5019 struct sd_lb_stats sds;
5021 init_sd_lb_stats(&sds);
5024 * Compute the various statistics relavent for load balancing at
5027 update_sd_lb_stats(env, &sds);
5028 local = &sds.local_stat;
5029 busiest = &sds.busiest_stat;
5031 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5032 check_asym_packing(env, &sds))
5035 /* There is no busy sibling group to pull tasks from */
5036 if (!sds.busiest || busiest->sum_nr_running == 0)
5039 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5042 * If the busiest group is imbalanced the below checks don't
5043 * work because they assume all things are equal, which typically
5044 * isn't true due to cpus_allowed constraints and the like.
5046 if (busiest->group_imb)
5049 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5050 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5051 !busiest->group_has_capacity)
5055 * If the local group is more busy than the selected busiest group
5056 * don't try and pull any tasks.
5058 if (local->avg_load >= busiest->avg_load)
5062 * Don't pull any tasks if this group is already above the domain
5065 if (local->avg_load >= sds.avg_load)
5068 if (env->idle == CPU_IDLE) {
5070 * This cpu is idle. If the busiest group load doesn't
5071 * have more tasks than the number of available cpu's and
5072 * there is no imbalance between this and busiest group
5073 * wrt to idle cpu's, it is balanced.
5075 if ((local->idle_cpus < busiest->idle_cpus) &&
5076 busiest->sum_nr_running <= busiest->group_weight)
5080 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5081 * imbalance_pct to be conservative.
5083 if (100 * busiest->avg_load <=
5084 env->sd->imbalance_pct * local->avg_load)
5089 /* Looks like there is an imbalance. Compute it */
5090 calculate_imbalance(env, &sds);
5099 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5101 static struct rq *find_busiest_queue(struct lb_env *env,
5102 struct sched_group *group)
5104 struct rq *busiest = NULL, *rq;
5105 unsigned long busiest_load = 0, busiest_power = 1;
5108 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5109 unsigned long power = power_of(i);
5110 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5115 capacity = fix_small_capacity(env->sd, group);
5118 wl = weighted_cpuload(i);
5121 * When comparing with imbalance, use weighted_cpuload()
5122 * which is not scaled with the cpu power.
5124 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5128 * For the load comparisons with the other cpu's, consider
5129 * the weighted_cpuload() scaled with the cpu power, so that
5130 * the load can be moved away from the cpu that is potentially
5131 * running at a lower capacity.
5133 * Thus we're looking for max(wl_i / power_i), crosswise
5134 * multiplication to rid ourselves of the division works out
5135 * to: wl_i * power_j > wl_j * power_i; where j is our
5138 if (wl * busiest_power > busiest_load * power) {
5140 busiest_power = power;
5149 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5150 * so long as it is large enough.
5152 #define MAX_PINNED_INTERVAL 512
5154 /* Working cpumask for load_balance and load_balance_newidle. */
5155 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5157 static int need_active_balance(struct lb_env *env)
5159 struct sched_domain *sd = env->sd;
5161 if (env->idle == CPU_NEWLY_IDLE) {
5164 * ASYM_PACKING needs to force migrate tasks from busy but
5165 * higher numbered CPUs in order to pack all tasks in the
5166 * lowest numbered CPUs.
5168 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5172 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5175 static int active_load_balance_cpu_stop(void *data);
5177 static int should_we_balance(struct lb_env *env)
5179 struct sched_group *sg = env->sd->groups;
5180 struct cpumask *sg_cpus, *sg_mask;
5181 int cpu, balance_cpu = -1;
5184 * In the newly idle case, we will allow all the cpu's
5185 * to do the newly idle load balance.
5187 if (env->idle == CPU_NEWLY_IDLE)
5190 sg_cpus = sched_group_cpus(sg);
5191 sg_mask = sched_group_mask(sg);
5192 /* Try to find first idle cpu */
5193 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5194 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5201 if (balance_cpu == -1)
5202 balance_cpu = group_balance_cpu(sg);
5205 * First idle cpu or the first cpu(busiest) in this sched group
5206 * is eligible for doing load balancing at this and above domains.
5208 return balance_cpu == env->dst_cpu;
5212 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5213 * tasks if there is an imbalance.
5215 static int load_balance(int this_cpu, struct rq *this_rq,
5216 struct sched_domain *sd, enum cpu_idle_type idle,
5217 int *continue_balancing)
5219 int ld_moved, cur_ld_moved, active_balance = 0;
5220 struct sched_domain *sd_parent = sd->parent;
5221 struct sched_group *group;
5223 unsigned long flags;
5224 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5226 struct lb_env env = {
5228 .dst_cpu = this_cpu,
5230 .dst_grpmask = sched_group_cpus(sd->groups),
5232 .loop_break = sched_nr_migrate_break,
5237 * For NEWLY_IDLE load_balancing, we don't need to consider
5238 * other cpus in our group
5240 if (idle == CPU_NEWLY_IDLE)
5241 env.dst_grpmask = NULL;
5243 cpumask_copy(cpus, cpu_active_mask);
5245 schedstat_inc(sd, lb_count[idle]);
5248 if (!should_we_balance(&env)) {
5249 *continue_balancing = 0;
5253 group = find_busiest_group(&env);
5255 schedstat_inc(sd, lb_nobusyg[idle]);
5259 busiest = find_busiest_queue(&env, group);
5261 schedstat_inc(sd, lb_nobusyq[idle]);
5265 BUG_ON(busiest == env.dst_rq);
5267 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5270 if (busiest->nr_running > 1) {
5272 * Attempt to move tasks. If find_busiest_group has found
5273 * an imbalance but busiest->nr_running <= 1, the group is
5274 * still unbalanced. ld_moved simply stays zero, so it is
5275 * correctly treated as an imbalance.
5277 env.flags |= LBF_ALL_PINNED;
5278 env.src_cpu = busiest->cpu;
5279 env.src_rq = busiest;
5280 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5283 local_irq_save(flags);
5284 double_rq_lock(env.dst_rq, busiest);
5287 * cur_ld_moved - load moved in current iteration
5288 * ld_moved - cumulative load moved across iterations
5290 cur_ld_moved = move_tasks(&env);
5291 ld_moved += cur_ld_moved;
5292 double_rq_unlock(env.dst_rq, busiest);
5293 local_irq_restore(flags);
5296 * some other cpu did the load balance for us.
5298 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5299 resched_cpu(env.dst_cpu);
5301 if (env.flags & LBF_NEED_BREAK) {
5302 env.flags &= ~LBF_NEED_BREAK;
5307 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5308 * us and move them to an alternate dst_cpu in our sched_group
5309 * where they can run. The upper limit on how many times we
5310 * iterate on same src_cpu is dependent on number of cpus in our
5313 * This changes load balance semantics a bit on who can move
5314 * load to a given_cpu. In addition to the given_cpu itself
5315 * (or a ilb_cpu acting on its behalf where given_cpu is
5316 * nohz-idle), we now have balance_cpu in a position to move
5317 * load to given_cpu. In rare situations, this may cause
5318 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5319 * _independently_ and at _same_ time to move some load to
5320 * given_cpu) causing exceess load to be moved to given_cpu.
5321 * This however should not happen so much in practice and
5322 * moreover subsequent load balance cycles should correct the
5323 * excess load moved.
5325 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5327 /* Prevent to re-select dst_cpu via env's cpus */
5328 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5330 env.dst_rq = cpu_rq(env.new_dst_cpu);
5331 env.dst_cpu = env.new_dst_cpu;
5332 env.flags &= ~LBF_DST_PINNED;
5334 env.loop_break = sched_nr_migrate_break;
5337 * Go back to "more_balance" rather than "redo" since we
5338 * need to continue with same src_cpu.
5344 * We failed to reach balance because of affinity.
5347 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5349 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5350 *group_imbalance = 1;
5351 } else if (*group_imbalance)
5352 *group_imbalance = 0;
5355 /* All tasks on this runqueue were pinned by CPU affinity */
5356 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5357 cpumask_clear_cpu(cpu_of(busiest), cpus);
5358 if (!cpumask_empty(cpus)) {
5360 env.loop_break = sched_nr_migrate_break;
5368 schedstat_inc(sd, lb_failed[idle]);
5370 * Increment the failure counter only on periodic balance.
5371 * We do not want newidle balance, which can be very
5372 * frequent, pollute the failure counter causing
5373 * excessive cache_hot migrations and active balances.
5375 if (idle != CPU_NEWLY_IDLE)
5376 sd->nr_balance_failed++;
5378 if (need_active_balance(&env)) {
5379 raw_spin_lock_irqsave(&busiest->lock, flags);
5381 /* don't kick the active_load_balance_cpu_stop,
5382 * if the curr task on busiest cpu can't be
5385 if (!cpumask_test_cpu(this_cpu,
5386 tsk_cpus_allowed(busiest->curr))) {
5387 raw_spin_unlock_irqrestore(&busiest->lock,
5389 env.flags |= LBF_ALL_PINNED;
5390 goto out_one_pinned;
5394 * ->active_balance synchronizes accesses to
5395 * ->active_balance_work. Once set, it's cleared
5396 * only after active load balance is finished.
5398 if (!busiest->active_balance) {
5399 busiest->active_balance = 1;
5400 busiest->push_cpu = this_cpu;
5403 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5405 if (active_balance) {
5406 stop_one_cpu_nowait(cpu_of(busiest),
5407 active_load_balance_cpu_stop, busiest,
5408 &busiest->active_balance_work);
5412 * We've kicked active balancing, reset the failure
5415 sd->nr_balance_failed = sd->cache_nice_tries+1;
5418 sd->nr_balance_failed = 0;
5420 if (likely(!active_balance)) {
5421 /* We were unbalanced, so reset the balancing interval */
5422 sd->balance_interval = sd->min_interval;
5425 * If we've begun active balancing, start to back off. This
5426 * case may not be covered by the all_pinned logic if there
5427 * is only 1 task on the busy runqueue (because we don't call
5430 if (sd->balance_interval < sd->max_interval)
5431 sd->balance_interval *= 2;
5437 schedstat_inc(sd, lb_balanced[idle]);
5439 sd->nr_balance_failed = 0;
5442 /* tune up the balancing interval */
5443 if (((env.flags & LBF_ALL_PINNED) &&
5444 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5445 (sd->balance_interval < sd->max_interval))
5446 sd->balance_interval *= 2;
5454 * idle_balance is called by schedule() if this_cpu is about to become
5455 * idle. Attempts to pull tasks from other CPUs.
5457 void idle_balance(int this_cpu, struct rq *this_rq)
5459 struct sched_domain *sd;
5460 int pulled_task = 0;
5461 unsigned long next_balance = jiffies + HZ;
5464 this_rq->idle_stamp = rq_clock(this_rq);
5466 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5470 * Drop the rq->lock, but keep IRQ/preempt disabled.
5472 raw_spin_unlock(&this_rq->lock);
5474 update_blocked_averages(this_cpu);
5476 for_each_domain(this_cpu, sd) {
5477 unsigned long interval;
5478 int continue_balancing = 1;
5479 u64 t0, domain_cost;
5481 if (!(sd->flags & SD_LOAD_BALANCE))
5484 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5487 if (sd->flags & SD_BALANCE_NEWIDLE) {
5488 t0 = sched_clock_cpu(this_cpu);
5490 /* If we've pulled tasks over stop searching: */
5491 pulled_task = load_balance(this_cpu, this_rq,
5493 &continue_balancing);
5495 domain_cost = sched_clock_cpu(this_cpu) - t0;
5496 if (domain_cost > sd->max_newidle_lb_cost)
5497 sd->max_newidle_lb_cost = domain_cost;
5499 curr_cost += domain_cost;
5502 interval = msecs_to_jiffies(sd->balance_interval);
5503 if (time_after(next_balance, sd->last_balance + interval))
5504 next_balance = sd->last_balance + interval;
5506 this_rq->idle_stamp = 0;
5512 raw_spin_lock(&this_rq->lock);
5514 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5516 * We are going idle. next_balance may be set based on
5517 * a busy processor. So reset next_balance.
5519 this_rq->next_balance = next_balance;
5522 if (curr_cost > this_rq->max_idle_balance_cost)
5523 this_rq->max_idle_balance_cost = curr_cost;
5527 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5528 * running tasks off the busiest CPU onto idle CPUs. It requires at
5529 * least 1 task to be running on each physical CPU where possible, and
5530 * avoids physical / logical imbalances.
5532 static int active_load_balance_cpu_stop(void *data)
5534 struct rq *busiest_rq = data;
5535 int busiest_cpu = cpu_of(busiest_rq);
5536 int target_cpu = busiest_rq->push_cpu;
5537 struct rq *target_rq = cpu_rq(target_cpu);
5538 struct sched_domain *sd;
5540 raw_spin_lock_irq(&busiest_rq->lock);
5542 /* make sure the requested cpu hasn't gone down in the meantime */
5543 if (unlikely(busiest_cpu != smp_processor_id() ||
5544 !busiest_rq->active_balance))
5547 /* Is there any task to move? */
5548 if (busiest_rq->nr_running <= 1)
5552 * This condition is "impossible", if it occurs
5553 * we need to fix it. Originally reported by
5554 * Bjorn Helgaas on a 128-cpu setup.
5556 BUG_ON(busiest_rq == target_rq);
5558 /* move a task from busiest_rq to target_rq */
5559 double_lock_balance(busiest_rq, target_rq);
5561 /* Search for an sd spanning us and the target CPU. */
5563 for_each_domain(target_cpu, sd) {
5564 if ((sd->flags & SD_LOAD_BALANCE) &&
5565 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5570 struct lb_env env = {
5572 .dst_cpu = target_cpu,
5573 .dst_rq = target_rq,
5574 .src_cpu = busiest_rq->cpu,
5575 .src_rq = busiest_rq,
5579 schedstat_inc(sd, alb_count);
5581 if (move_one_task(&env))
5582 schedstat_inc(sd, alb_pushed);
5584 schedstat_inc(sd, alb_failed);
5587 double_unlock_balance(busiest_rq, target_rq);
5589 busiest_rq->active_balance = 0;
5590 raw_spin_unlock_irq(&busiest_rq->lock);
5594 #ifdef CONFIG_NO_HZ_COMMON
5596 * idle load balancing details
5597 * - When one of the busy CPUs notice that there may be an idle rebalancing
5598 * needed, they will kick the idle load balancer, which then does idle
5599 * load balancing for all the idle CPUs.
5602 cpumask_var_t idle_cpus_mask;
5604 unsigned long next_balance; /* in jiffy units */
5605 } nohz ____cacheline_aligned;
5607 static inline int find_new_ilb(int call_cpu)
5609 int ilb = cpumask_first(nohz.idle_cpus_mask);
5611 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5618 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5619 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5620 * CPU (if there is one).
5622 static void nohz_balancer_kick(int cpu)
5626 nohz.next_balance++;
5628 ilb_cpu = find_new_ilb(cpu);
5630 if (ilb_cpu >= nr_cpu_ids)
5633 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5636 * Use smp_send_reschedule() instead of resched_cpu().
5637 * This way we generate a sched IPI on the target cpu which
5638 * is idle. And the softirq performing nohz idle load balance
5639 * will be run before returning from the IPI.
5641 smp_send_reschedule(ilb_cpu);
5645 static inline void nohz_balance_exit_idle(int cpu)
5647 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5648 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5649 atomic_dec(&nohz.nr_cpus);
5650 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5654 static inline void set_cpu_sd_state_busy(void)
5656 struct sched_domain *sd;
5659 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5661 if (!sd || !sd->nohz_idle)
5665 for (; sd; sd = sd->parent)
5666 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5671 void set_cpu_sd_state_idle(void)
5673 struct sched_domain *sd;
5676 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5678 if (!sd || sd->nohz_idle)
5682 for (; sd; sd = sd->parent)
5683 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5689 * This routine will record that the cpu is going idle with tick stopped.
5690 * This info will be used in performing idle load balancing in the future.
5692 void nohz_balance_enter_idle(int cpu)
5695 * If this cpu is going down, then nothing needs to be done.
5697 if (!cpu_active(cpu))
5700 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5703 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5704 atomic_inc(&nohz.nr_cpus);
5705 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5708 static int sched_ilb_notifier(struct notifier_block *nfb,
5709 unsigned long action, void *hcpu)
5711 switch (action & ~CPU_TASKS_FROZEN) {
5713 nohz_balance_exit_idle(smp_processor_id());
5721 static DEFINE_SPINLOCK(balancing);
5724 * Scale the max load_balance interval with the number of CPUs in the system.
5725 * This trades load-balance latency on larger machines for less cross talk.
5727 void update_max_interval(void)
5729 max_load_balance_interval = HZ*num_online_cpus()/10;
5733 * It checks each scheduling domain to see if it is due to be balanced,
5734 * and initiates a balancing operation if so.
5736 * Balancing parameters are set up in init_sched_domains.
5738 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5740 int continue_balancing = 1;
5741 struct rq *rq = cpu_rq(cpu);
5742 unsigned long interval;
5743 struct sched_domain *sd;
5744 /* Earliest time when we have to do rebalance again */
5745 unsigned long next_balance = jiffies + 60*HZ;
5746 int update_next_balance = 0;
5747 int need_serialize, need_decay = 0;
5750 update_blocked_averages(cpu);
5753 for_each_domain(cpu, sd) {
5755 * Decay the newidle max times here because this is a regular
5756 * visit to all the domains. Decay ~1% per second.
5758 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5759 sd->max_newidle_lb_cost =
5760 (sd->max_newidle_lb_cost * 253) / 256;
5761 sd->next_decay_max_lb_cost = jiffies + HZ;
5764 max_cost += sd->max_newidle_lb_cost;
5766 if (!(sd->flags & SD_LOAD_BALANCE))
5770 * Stop the load balance at this level. There is another
5771 * CPU in our sched group which is doing load balancing more
5774 if (!continue_balancing) {
5780 interval = sd->balance_interval;
5781 if (idle != CPU_IDLE)
5782 interval *= sd->busy_factor;
5784 /* scale ms to jiffies */
5785 interval = msecs_to_jiffies(interval);
5786 interval = clamp(interval, 1UL, max_load_balance_interval);
5788 need_serialize = sd->flags & SD_SERIALIZE;
5790 if (need_serialize) {
5791 if (!spin_trylock(&balancing))
5795 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5796 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5798 * The LBF_DST_PINNED logic could have changed
5799 * env->dst_cpu, so we can't know our idle
5800 * state even if we migrated tasks. Update it.
5802 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5804 sd->last_balance = jiffies;
5807 spin_unlock(&balancing);
5809 if (time_after(next_balance, sd->last_balance + interval)) {
5810 next_balance = sd->last_balance + interval;
5811 update_next_balance = 1;
5816 * Ensure the rq-wide value also decays but keep it at a
5817 * reasonable floor to avoid funnies with rq->avg_idle.
5819 rq->max_idle_balance_cost =
5820 max((u64)sysctl_sched_migration_cost, max_cost);
5825 * next_balance will be updated only when there is a need.
5826 * When the cpu is attached to null domain for ex, it will not be
5829 if (likely(update_next_balance))
5830 rq->next_balance = next_balance;
5833 #ifdef CONFIG_NO_HZ_COMMON
5835 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5836 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5838 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5840 struct rq *this_rq = cpu_rq(this_cpu);
5844 if (idle != CPU_IDLE ||
5845 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5848 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5849 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5853 * If this cpu gets work to do, stop the load balancing
5854 * work being done for other cpus. Next load
5855 * balancing owner will pick it up.
5860 rq = cpu_rq(balance_cpu);
5862 raw_spin_lock_irq(&rq->lock);
5863 update_rq_clock(rq);
5864 update_idle_cpu_load(rq);
5865 raw_spin_unlock_irq(&rq->lock);
5867 rebalance_domains(balance_cpu, CPU_IDLE);
5869 if (time_after(this_rq->next_balance, rq->next_balance))
5870 this_rq->next_balance = rq->next_balance;
5872 nohz.next_balance = this_rq->next_balance;
5874 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5878 * Current heuristic for kicking the idle load balancer in the presence
5879 * of an idle cpu is the system.
5880 * - This rq has more than one task.
5881 * - At any scheduler domain level, this cpu's scheduler group has multiple
5882 * busy cpu's exceeding the group's power.
5883 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5884 * domain span are idle.
5886 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5888 unsigned long now = jiffies;
5889 struct sched_domain *sd;
5891 if (unlikely(idle_cpu(cpu)))
5895 * We may be recently in ticked or tickless idle mode. At the first
5896 * busy tick after returning from idle, we will update the busy stats.
5898 set_cpu_sd_state_busy();
5899 nohz_balance_exit_idle(cpu);
5902 * None are in tickless mode and hence no need for NOHZ idle load
5905 if (likely(!atomic_read(&nohz.nr_cpus)))
5908 if (time_before(now, nohz.next_balance))
5911 if (rq->nr_running >= 2)
5915 for_each_domain(cpu, sd) {
5916 struct sched_group *sg = sd->groups;
5917 struct sched_group_power *sgp = sg->sgp;
5918 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5920 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5921 goto need_kick_unlock;
5923 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5924 && (cpumask_first_and(nohz.idle_cpus_mask,
5925 sched_domain_span(sd)) < cpu))
5926 goto need_kick_unlock;
5928 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5940 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5944 * run_rebalance_domains is triggered when needed from the scheduler tick.
5945 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5947 static void run_rebalance_domains(struct softirq_action *h)
5949 int this_cpu = smp_processor_id();
5950 struct rq *this_rq = cpu_rq(this_cpu);
5951 enum cpu_idle_type idle = this_rq->idle_balance ?
5952 CPU_IDLE : CPU_NOT_IDLE;
5954 rebalance_domains(this_cpu, idle);
5957 * If this cpu has a pending nohz_balance_kick, then do the
5958 * balancing on behalf of the other idle cpus whose ticks are
5961 nohz_idle_balance(this_cpu, idle);
5964 static inline int on_null_domain(int cpu)
5966 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5970 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5972 void trigger_load_balance(struct rq *rq, int cpu)
5974 /* Don't need to rebalance while attached to NULL domain */
5975 if (time_after_eq(jiffies, rq->next_balance) &&
5976 likely(!on_null_domain(cpu)))
5977 raise_softirq(SCHED_SOFTIRQ);
5978 #ifdef CONFIG_NO_HZ_COMMON
5979 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5980 nohz_balancer_kick(cpu);
5984 static void rq_online_fair(struct rq *rq)
5989 static void rq_offline_fair(struct rq *rq)
5993 /* Ensure any throttled groups are reachable by pick_next_task */
5994 unthrottle_offline_cfs_rqs(rq);
5997 #endif /* CONFIG_SMP */
6000 * scheduler tick hitting a task of our scheduling class:
6002 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6004 struct cfs_rq *cfs_rq;
6005 struct sched_entity *se = &curr->se;
6007 for_each_sched_entity(se) {
6008 cfs_rq = cfs_rq_of(se);
6009 entity_tick(cfs_rq, se, queued);
6012 if (numabalancing_enabled)
6013 task_tick_numa(rq, curr);
6015 update_rq_runnable_avg(rq, 1);
6019 * called on fork with the child task as argument from the parent's context
6020 * - child not yet on the tasklist
6021 * - preemption disabled
6023 static void task_fork_fair(struct task_struct *p)
6025 struct cfs_rq *cfs_rq;
6026 struct sched_entity *se = &p->se, *curr;
6027 int this_cpu = smp_processor_id();
6028 struct rq *rq = this_rq();
6029 unsigned long flags;
6031 raw_spin_lock_irqsave(&rq->lock, flags);
6033 update_rq_clock(rq);
6035 cfs_rq = task_cfs_rq(current);
6036 curr = cfs_rq->curr;
6039 * Not only the cpu but also the task_group of the parent might have
6040 * been changed after parent->se.parent,cfs_rq were copied to
6041 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6042 * of child point to valid ones.
6045 __set_task_cpu(p, this_cpu);
6048 update_curr(cfs_rq);
6051 se->vruntime = curr->vruntime;
6052 place_entity(cfs_rq, se, 1);
6054 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6056 * Upon rescheduling, sched_class::put_prev_task() will place
6057 * 'current' within the tree based on its new key value.
6059 swap(curr->vruntime, se->vruntime);
6060 resched_task(rq->curr);
6063 se->vruntime -= cfs_rq->min_vruntime;
6065 raw_spin_unlock_irqrestore(&rq->lock, flags);
6069 * Priority of the task has changed. Check to see if we preempt
6073 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6079 * Reschedule if we are currently running on this runqueue and
6080 * our priority decreased, or if we are not currently running on
6081 * this runqueue and our priority is higher than the current's
6083 if (rq->curr == p) {
6084 if (p->prio > oldprio)
6085 resched_task(rq->curr);
6087 check_preempt_curr(rq, p, 0);
6090 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6092 struct sched_entity *se = &p->se;
6093 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6096 * Ensure the task's vruntime is normalized, so that when its
6097 * switched back to the fair class the enqueue_entity(.flags=0) will
6098 * do the right thing.
6100 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6101 * have normalized the vruntime, if it was !on_rq, then only when
6102 * the task is sleeping will it still have non-normalized vruntime.
6104 if (!se->on_rq && p->state != TASK_RUNNING) {
6106 * Fix up our vruntime so that the current sleep doesn't
6107 * cause 'unlimited' sleep bonus.
6109 place_entity(cfs_rq, se, 0);
6110 se->vruntime -= cfs_rq->min_vruntime;
6115 * Remove our load from contribution when we leave sched_fair
6116 * and ensure we don't carry in an old decay_count if we
6119 if (se->avg.decay_count) {
6120 __synchronize_entity_decay(se);
6121 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6127 * We switched to the sched_fair class.
6129 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6135 * We were most likely switched from sched_rt, so
6136 * kick off the schedule if running, otherwise just see
6137 * if we can still preempt the current task.
6140 resched_task(rq->curr);
6142 check_preempt_curr(rq, p, 0);
6145 /* Account for a task changing its policy or group.
6147 * This routine is mostly called to set cfs_rq->curr field when a task
6148 * migrates between groups/classes.
6150 static void set_curr_task_fair(struct rq *rq)
6152 struct sched_entity *se = &rq->curr->se;
6154 for_each_sched_entity(se) {
6155 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6157 set_next_entity(cfs_rq, se);
6158 /* ensure bandwidth has been allocated on our new cfs_rq */
6159 account_cfs_rq_runtime(cfs_rq, 0);
6163 void init_cfs_rq(struct cfs_rq *cfs_rq)
6165 cfs_rq->tasks_timeline = RB_ROOT;
6166 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6167 #ifndef CONFIG_64BIT
6168 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6171 atomic64_set(&cfs_rq->decay_counter, 1);
6172 atomic_long_set(&cfs_rq->removed_load, 0);
6176 #ifdef CONFIG_FAIR_GROUP_SCHED
6177 static void task_move_group_fair(struct task_struct *p, int on_rq)
6179 struct cfs_rq *cfs_rq;
6181 * If the task was not on the rq at the time of this cgroup movement
6182 * it must have been asleep, sleeping tasks keep their ->vruntime
6183 * absolute on their old rq until wakeup (needed for the fair sleeper
6184 * bonus in place_entity()).
6186 * If it was on the rq, we've just 'preempted' it, which does convert
6187 * ->vruntime to a relative base.
6189 * Make sure both cases convert their relative position when migrating
6190 * to another cgroup's rq. This does somewhat interfere with the
6191 * fair sleeper stuff for the first placement, but who cares.
6194 * When !on_rq, vruntime of the task has usually NOT been normalized.
6195 * But there are some cases where it has already been normalized:
6197 * - Moving a forked child which is waiting for being woken up by
6198 * wake_up_new_task().
6199 * - Moving a task which has been woken up by try_to_wake_up() and
6200 * waiting for actually being woken up by sched_ttwu_pending().
6202 * To prevent boost or penalty in the new cfs_rq caused by delta
6203 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6205 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6209 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6210 set_task_rq(p, task_cpu(p));
6212 cfs_rq = cfs_rq_of(&p->se);
6213 p->se.vruntime += cfs_rq->min_vruntime;
6216 * migrate_task_rq_fair() will have removed our previous
6217 * contribution, but we must synchronize for ongoing future
6220 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6221 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6226 void free_fair_sched_group(struct task_group *tg)
6230 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6232 for_each_possible_cpu(i) {
6234 kfree(tg->cfs_rq[i]);
6243 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6245 struct cfs_rq *cfs_rq;
6246 struct sched_entity *se;
6249 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6252 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6256 tg->shares = NICE_0_LOAD;
6258 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6260 for_each_possible_cpu(i) {
6261 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6262 GFP_KERNEL, cpu_to_node(i));
6266 se = kzalloc_node(sizeof(struct sched_entity),
6267 GFP_KERNEL, cpu_to_node(i));
6271 init_cfs_rq(cfs_rq);
6272 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6283 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6285 struct rq *rq = cpu_rq(cpu);
6286 unsigned long flags;
6289 * Only empty task groups can be destroyed; so we can speculatively
6290 * check on_list without danger of it being re-added.
6292 if (!tg->cfs_rq[cpu]->on_list)
6295 raw_spin_lock_irqsave(&rq->lock, flags);
6296 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6297 raw_spin_unlock_irqrestore(&rq->lock, flags);
6300 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6301 struct sched_entity *se, int cpu,
6302 struct sched_entity *parent)
6304 struct rq *rq = cpu_rq(cpu);
6308 init_cfs_rq_runtime(cfs_rq);
6310 tg->cfs_rq[cpu] = cfs_rq;
6313 /* se could be NULL for root_task_group */
6318 se->cfs_rq = &rq->cfs;
6320 se->cfs_rq = parent->my_q;
6323 update_load_set(&se->load, 0);
6324 se->parent = parent;
6327 static DEFINE_MUTEX(shares_mutex);
6329 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6332 unsigned long flags;
6335 * We can't change the weight of the root cgroup.
6340 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6342 mutex_lock(&shares_mutex);
6343 if (tg->shares == shares)
6346 tg->shares = shares;
6347 for_each_possible_cpu(i) {
6348 struct rq *rq = cpu_rq(i);
6349 struct sched_entity *se;
6352 /* Propagate contribution to hierarchy */
6353 raw_spin_lock_irqsave(&rq->lock, flags);
6355 /* Possible calls to update_curr() need rq clock */
6356 update_rq_clock(rq);
6357 for_each_sched_entity(se)
6358 update_cfs_shares(group_cfs_rq(se));
6359 raw_spin_unlock_irqrestore(&rq->lock, flags);
6363 mutex_unlock(&shares_mutex);
6366 #else /* CONFIG_FAIR_GROUP_SCHED */
6368 void free_fair_sched_group(struct task_group *tg) { }
6370 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6375 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6377 #endif /* CONFIG_FAIR_GROUP_SCHED */
6380 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6382 struct sched_entity *se = &task->se;
6383 unsigned int rr_interval = 0;
6386 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6389 if (rq->cfs.load.weight)
6390 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6396 * All the scheduling class methods:
6398 const struct sched_class fair_sched_class = {
6399 .next = &idle_sched_class,
6400 .enqueue_task = enqueue_task_fair,
6401 .dequeue_task = dequeue_task_fair,
6402 .yield_task = yield_task_fair,
6403 .yield_to_task = yield_to_task_fair,
6405 .check_preempt_curr = check_preempt_wakeup,
6407 .pick_next_task = pick_next_task_fair,
6408 .put_prev_task = put_prev_task_fair,
6411 .select_task_rq = select_task_rq_fair,
6412 .migrate_task_rq = migrate_task_rq_fair,
6414 .rq_online = rq_online_fair,
6415 .rq_offline = rq_offline_fair,
6417 .task_waking = task_waking_fair,
6420 .set_curr_task = set_curr_task_fair,
6421 .task_tick = task_tick_fair,
6422 .task_fork = task_fork_fair,
6424 .prio_changed = prio_changed_fair,
6425 .switched_from = switched_from_fair,
6426 .switched_to = switched_to_fair,
6428 .get_rr_interval = get_rr_interval_fair,
6430 #ifdef CONFIG_FAIR_GROUP_SCHED
6431 .task_move_group = task_move_group_fair,
6435 #ifdef CONFIG_SCHED_DEBUG
6436 void print_cfs_stats(struct seq_file *m, int cpu)
6438 struct cfs_rq *cfs_rq;
6441 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6442 print_cfs_rq(m, cpu, cfs_rq);
6447 __init void init_sched_fair_class(void)
6450 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6452 #ifdef CONFIG_NO_HZ_COMMON
6453 nohz.next_balance = jiffies;
6454 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6455 cpu_notifier(sched_ilb_notifier, 0);