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 * It is possible to reach the end of the VMA list but the last few
1043 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1044 * would find the !migratable VMA on the next scan but not reset the
1045 * scanner to the start so check it now.
1048 mm->numa_scan_offset = start;
1050 reset_ptenuma_scan(p);
1051 up_read(&mm->mmap_sem);
1055 * Drive the periodic memory faults..
1057 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1059 struct callback_head *work = &curr->numa_work;
1063 * We don't care about NUMA placement if we don't have memory.
1065 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1069 * Using runtime rather than walltime has the dual advantage that
1070 * we (mostly) drive the selection from busy threads and that the
1071 * task needs to have done some actual work before we bother with
1074 now = curr->se.sum_exec_runtime;
1075 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1077 if (now - curr->node_stamp > period) {
1078 if (!curr->node_stamp)
1079 curr->numa_scan_period = task_scan_min(curr);
1080 curr->node_stamp += period;
1082 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1083 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1084 task_work_add(curr, work, true);
1089 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1092 #endif /* CONFIG_NUMA_BALANCING */
1095 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1097 update_load_add(&cfs_rq->load, se->load.weight);
1098 if (!parent_entity(se))
1099 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1101 if (entity_is_task(se))
1102 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1104 cfs_rq->nr_running++;
1108 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1110 update_load_sub(&cfs_rq->load, se->load.weight);
1111 if (!parent_entity(se))
1112 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1113 if (entity_is_task(se))
1114 list_del_init(&se->group_node);
1115 cfs_rq->nr_running--;
1118 #ifdef CONFIG_FAIR_GROUP_SCHED
1120 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1125 * Use this CPU's actual weight instead of the last load_contribution
1126 * to gain a more accurate current total weight. See
1127 * update_cfs_rq_load_contribution().
1129 tg_weight = atomic_long_read(&tg->load_avg);
1130 tg_weight -= cfs_rq->tg_load_contrib;
1131 tg_weight += cfs_rq->load.weight;
1136 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1138 long tg_weight, load, shares;
1140 tg_weight = calc_tg_weight(tg, cfs_rq);
1141 load = cfs_rq->load.weight;
1143 shares = (tg->shares * load);
1145 shares /= tg_weight;
1147 if (shares < MIN_SHARES)
1148 shares = MIN_SHARES;
1149 if (shares > tg->shares)
1150 shares = tg->shares;
1154 # else /* CONFIG_SMP */
1155 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1159 # endif /* CONFIG_SMP */
1160 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1161 unsigned long weight)
1164 /* commit outstanding execution time */
1165 if (cfs_rq->curr == se)
1166 update_curr(cfs_rq);
1167 account_entity_dequeue(cfs_rq, se);
1170 update_load_set(&se->load, weight);
1173 account_entity_enqueue(cfs_rq, se);
1176 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1178 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1180 struct task_group *tg;
1181 struct sched_entity *se;
1185 se = tg->se[cpu_of(rq_of(cfs_rq))];
1186 if (!se || throttled_hierarchy(cfs_rq))
1189 if (likely(se->load.weight == tg->shares))
1192 shares = calc_cfs_shares(cfs_rq, tg);
1194 reweight_entity(cfs_rq_of(se), se, shares);
1196 #else /* CONFIG_FAIR_GROUP_SCHED */
1197 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1200 #endif /* CONFIG_FAIR_GROUP_SCHED */
1204 * We choose a half-life close to 1 scheduling period.
1205 * Note: The tables below are dependent on this value.
1207 #define LOAD_AVG_PERIOD 32
1208 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1209 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1211 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1212 static const u32 runnable_avg_yN_inv[] = {
1213 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1214 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1215 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1216 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1217 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1218 0x85aac367, 0x82cd8698,
1222 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1223 * over-estimates when re-combining.
1225 static const u32 runnable_avg_yN_sum[] = {
1226 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1227 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1228 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1233 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1235 static __always_inline u64 decay_load(u64 val, u64 n)
1237 unsigned int local_n;
1241 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1244 /* after bounds checking we can collapse to 32-bit */
1248 * As y^PERIOD = 1/2, we can combine
1249 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1250 * With a look-up table which covers k^n (n<PERIOD)
1252 * To achieve constant time decay_load.
1254 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1255 val >>= local_n / LOAD_AVG_PERIOD;
1256 local_n %= LOAD_AVG_PERIOD;
1259 val *= runnable_avg_yN_inv[local_n];
1260 /* We don't use SRR here since we always want to round down. */
1265 * For updates fully spanning n periods, the contribution to runnable
1266 * average will be: \Sum 1024*y^n
1268 * We can compute this reasonably efficiently by combining:
1269 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1271 static u32 __compute_runnable_contrib(u64 n)
1275 if (likely(n <= LOAD_AVG_PERIOD))
1276 return runnable_avg_yN_sum[n];
1277 else if (unlikely(n >= LOAD_AVG_MAX_N))
1278 return LOAD_AVG_MAX;
1280 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1282 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1283 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1285 n -= LOAD_AVG_PERIOD;
1286 } while (n > LOAD_AVG_PERIOD);
1288 contrib = decay_load(contrib, n);
1289 return contrib + runnable_avg_yN_sum[n];
1293 * We can represent the historical contribution to runnable average as the
1294 * coefficients of a geometric series. To do this we sub-divide our runnable
1295 * history into segments of approximately 1ms (1024us); label the segment that
1296 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1298 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1300 * (now) (~1ms ago) (~2ms ago)
1302 * Let u_i denote the fraction of p_i that the entity was runnable.
1304 * We then designate the fractions u_i as our co-efficients, yielding the
1305 * following representation of historical load:
1306 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1308 * We choose y based on the with of a reasonably scheduling period, fixing:
1311 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1312 * approximately half as much as the contribution to load within the last ms
1315 * When a period "rolls over" and we have new u_0`, multiplying the previous
1316 * sum again by y is sufficient to update:
1317 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1318 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1320 static __always_inline int __update_entity_runnable_avg(u64 now,
1321 struct sched_avg *sa,
1325 u32 runnable_contrib;
1326 int delta_w, decayed = 0;
1328 delta = now - sa->last_runnable_update;
1330 * This should only happen when time goes backwards, which it
1331 * unfortunately does during sched clock init when we swap over to TSC.
1333 if ((s64)delta < 0) {
1334 sa->last_runnable_update = now;
1339 * Use 1024ns as the unit of measurement since it's a reasonable
1340 * approximation of 1us and fast to compute.
1345 sa->last_runnable_update = now;
1347 /* delta_w is the amount already accumulated against our next period */
1348 delta_w = sa->runnable_avg_period % 1024;
1349 if (delta + delta_w >= 1024) {
1350 /* period roll-over */
1354 * Now that we know we're crossing a period boundary, figure
1355 * out how much from delta we need to complete the current
1356 * period and accrue it.
1358 delta_w = 1024 - delta_w;
1360 sa->runnable_avg_sum += delta_w;
1361 sa->runnable_avg_period += delta_w;
1365 /* Figure out how many additional periods this update spans */
1366 periods = delta / 1024;
1369 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1371 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1374 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1375 runnable_contrib = __compute_runnable_contrib(periods);
1377 sa->runnable_avg_sum += runnable_contrib;
1378 sa->runnable_avg_period += runnable_contrib;
1381 /* Remainder of delta accrued against u_0` */
1383 sa->runnable_avg_sum += delta;
1384 sa->runnable_avg_period += delta;
1389 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1390 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1392 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1393 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1395 decays -= se->avg.decay_count;
1399 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1400 se->avg.decay_count = 0;
1405 #ifdef CONFIG_FAIR_GROUP_SCHED
1406 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1409 struct task_group *tg = cfs_rq->tg;
1412 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1413 tg_contrib -= cfs_rq->tg_load_contrib;
1415 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1416 atomic_long_add(tg_contrib, &tg->load_avg);
1417 cfs_rq->tg_load_contrib += tg_contrib;
1422 * Aggregate cfs_rq runnable averages into an equivalent task_group
1423 * representation for computing load contributions.
1425 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1426 struct cfs_rq *cfs_rq)
1428 struct task_group *tg = cfs_rq->tg;
1431 /* The fraction of a cpu used by this cfs_rq */
1432 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1433 sa->runnable_avg_period + 1);
1434 contrib -= cfs_rq->tg_runnable_contrib;
1436 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1437 atomic_add(contrib, &tg->runnable_avg);
1438 cfs_rq->tg_runnable_contrib += contrib;
1442 static inline void __update_group_entity_contrib(struct sched_entity *se)
1444 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1445 struct task_group *tg = cfs_rq->tg;
1450 contrib = cfs_rq->tg_load_contrib * tg->shares;
1451 se->avg.load_avg_contrib = div_u64(contrib,
1452 atomic_long_read(&tg->load_avg) + 1);
1455 * For group entities we need to compute a correction term in the case
1456 * that they are consuming <1 cpu so that we would contribute the same
1457 * load as a task of equal weight.
1459 * Explicitly co-ordinating this measurement would be expensive, but
1460 * fortunately the sum of each cpus contribution forms a usable
1461 * lower-bound on the true value.
1463 * Consider the aggregate of 2 contributions. Either they are disjoint
1464 * (and the sum represents true value) or they are disjoint and we are
1465 * understating by the aggregate of their overlap.
1467 * Extending this to N cpus, for a given overlap, the maximum amount we
1468 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1469 * cpus that overlap for this interval and w_i is the interval width.
1471 * On a small machine; the first term is well-bounded which bounds the
1472 * total error since w_i is a subset of the period. Whereas on a
1473 * larger machine, while this first term can be larger, if w_i is the
1474 * of consequential size guaranteed to see n_i*w_i quickly converge to
1475 * our upper bound of 1-cpu.
1477 runnable_avg = atomic_read(&tg->runnable_avg);
1478 if (runnable_avg < NICE_0_LOAD) {
1479 se->avg.load_avg_contrib *= runnable_avg;
1480 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1484 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1485 int force_update) {}
1486 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1487 struct cfs_rq *cfs_rq) {}
1488 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1491 static inline void __update_task_entity_contrib(struct sched_entity *se)
1495 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1496 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1497 contrib /= (se->avg.runnable_avg_period + 1);
1498 se->avg.load_avg_contrib = scale_load(contrib);
1501 /* Compute the current contribution to load_avg by se, return any delta */
1502 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1504 long old_contrib = se->avg.load_avg_contrib;
1506 if (entity_is_task(se)) {
1507 __update_task_entity_contrib(se);
1509 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1510 __update_group_entity_contrib(se);
1513 return se->avg.load_avg_contrib - old_contrib;
1516 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1519 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1520 cfs_rq->blocked_load_avg -= load_contrib;
1522 cfs_rq->blocked_load_avg = 0;
1525 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1527 /* Update a sched_entity's runnable average */
1528 static inline void update_entity_load_avg(struct sched_entity *se,
1531 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1536 * For a group entity we need to use their owned cfs_rq_clock_task() in
1537 * case they are the parent of a throttled hierarchy.
1539 if (entity_is_task(se))
1540 now = cfs_rq_clock_task(cfs_rq);
1542 now = cfs_rq_clock_task(group_cfs_rq(se));
1544 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1547 contrib_delta = __update_entity_load_avg_contrib(se);
1553 cfs_rq->runnable_load_avg += contrib_delta;
1555 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1559 * Decay the load contributed by all blocked children and account this so that
1560 * their contribution may appropriately discounted when they wake up.
1562 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1564 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1567 decays = now - cfs_rq->last_decay;
1568 if (!decays && !force_update)
1571 if (atomic_long_read(&cfs_rq->removed_load)) {
1572 unsigned long removed_load;
1573 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1574 subtract_blocked_load_contrib(cfs_rq, removed_load);
1578 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1580 atomic64_add(decays, &cfs_rq->decay_counter);
1581 cfs_rq->last_decay = now;
1584 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1587 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1589 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1590 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1593 /* Add the load generated by se into cfs_rq's child load-average */
1594 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1595 struct sched_entity *se,
1599 * We track migrations using entity decay_count <= 0, on a wake-up
1600 * migration we use a negative decay count to track the remote decays
1601 * accumulated while sleeping.
1603 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1604 * are seen by enqueue_entity_load_avg() as a migration with an already
1605 * constructed load_avg_contrib.
1607 if (unlikely(se->avg.decay_count <= 0)) {
1608 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1609 if (se->avg.decay_count) {
1611 * In a wake-up migration we have to approximate the
1612 * time sleeping. This is because we can't synchronize
1613 * clock_task between the two cpus, and it is not
1614 * guaranteed to be read-safe. Instead, we can
1615 * approximate this using our carried decays, which are
1616 * explicitly atomically readable.
1618 se->avg.last_runnable_update -= (-se->avg.decay_count)
1620 update_entity_load_avg(se, 0);
1621 /* Indicate that we're now synchronized and on-rq */
1622 se->avg.decay_count = 0;
1627 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1628 * would have made count negative); we must be careful to avoid
1629 * double-accounting blocked time after synchronizing decays.
1631 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1635 /* migrated tasks did not contribute to our blocked load */
1637 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1638 update_entity_load_avg(se, 0);
1641 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1642 /* we force update consideration on load-balancer moves */
1643 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1647 * Remove se's load from this cfs_rq child load-average, if the entity is
1648 * transitioning to a blocked state we track its projected decay using
1651 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1652 struct sched_entity *se,
1655 update_entity_load_avg(se, 1);
1656 /* we force update consideration on load-balancer moves */
1657 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1659 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1661 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1662 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1663 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1667 * Update the rq's load with the elapsed running time before entering
1668 * idle. if the last scheduled task is not a CFS task, idle_enter will
1669 * be the only way to update the runnable statistic.
1671 void idle_enter_fair(struct rq *this_rq)
1673 update_rq_runnable_avg(this_rq, 1);
1677 * Update the rq's load with the elapsed idle time before a task is
1678 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1679 * be the only way to update the runnable statistic.
1681 void idle_exit_fair(struct rq *this_rq)
1683 update_rq_runnable_avg(this_rq, 0);
1687 static inline void update_entity_load_avg(struct sched_entity *se,
1688 int update_cfs_rq) {}
1689 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1690 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1691 struct sched_entity *se,
1693 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1694 struct sched_entity *se,
1696 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1697 int force_update) {}
1700 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1702 #ifdef CONFIG_SCHEDSTATS
1703 struct task_struct *tsk = NULL;
1705 if (entity_is_task(se))
1708 if (se->statistics.sleep_start) {
1709 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1714 if (unlikely(delta > se->statistics.sleep_max))
1715 se->statistics.sleep_max = delta;
1717 se->statistics.sleep_start = 0;
1718 se->statistics.sum_sleep_runtime += delta;
1721 account_scheduler_latency(tsk, delta >> 10, 1);
1722 trace_sched_stat_sleep(tsk, delta);
1725 if (se->statistics.block_start) {
1726 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1731 if (unlikely(delta > se->statistics.block_max))
1732 se->statistics.block_max = delta;
1734 se->statistics.block_start = 0;
1735 se->statistics.sum_sleep_runtime += delta;
1738 if (tsk->in_iowait) {
1739 se->statistics.iowait_sum += delta;
1740 se->statistics.iowait_count++;
1741 trace_sched_stat_iowait(tsk, delta);
1744 trace_sched_stat_blocked(tsk, delta);
1747 * Blocking time is in units of nanosecs, so shift by
1748 * 20 to get a milliseconds-range estimation of the
1749 * amount of time that the task spent sleeping:
1751 if (unlikely(prof_on == SLEEP_PROFILING)) {
1752 profile_hits(SLEEP_PROFILING,
1753 (void *)get_wchan(tsk),
1756 account_scheduler_latency(tsk, delta >> 10, 0);
1762 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1764 #ifdef CONFIG_SCHED_DEBUG
1765 s64 d = se->vruntime - cfs_rq->min_vruntime;
1770 if (d > 3*sysctl_sched_latency)
1771 schedstat_inc(cfs_rq, nr_spread_over);
1776 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1778 u64 vruntime = cfs_rq->min_vruntime;
1781 * The 'current' period is already promised to the current tasks,
1782 * however the extra weight of the new task will slow them down a
1783 * little, place the new task so that it fits in the slot that
1784 * stays open at the end.
1786 if (initial && sched_feat(START_DEBIT))
1787 vruntime += sched_vslice(cfs_rq, se);
1789 /* sleeps up to a single latency don't count. */
1791 unsigned long thresh = sysctl_sched_latency;
1794 * Halve their sleep time's effect, to allow
1795 * for a gentler effect of sleepers:
1797 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1803 /* ensure we never gain time by being placed backwards. */
1804 se->vruntime = max_vruntime(se->vruntime, vruntime);
1807 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1810 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1813 * Update the normalized vruntime before updating min_vruntime
1814 * through calling update_curr().
1816 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1817 se->vruntime += cfs_rq->min_vruntime;
1820 * Update run-time statistics of the 'current'.
1822 update_curr(cfs_rq);
1823 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1824 account_entity_enqueue(cfs_rq, se);
1825 update_cfs_shares(cfs_rq);
1827 if (flags & ENQUEUE_WAKEUP) {
1828 place_entity(cfs_rq, se, 0);
1829 enqueue_sleeper(cfs_rq, se);
1832 update_stats_enqueue(cfs_rq, se);
1833 check_spread(cfs_rq, se);
1834 if (se != cfs_rq->curr)
1835 __enqueue_entity(cfs_rq, se);
1838 if (cfs_rq->nr_running == 1) {
1839 list_add_leaf_cfs_rq(cfs_rq);
1840 check_enqueue_throttle(cfs_rq);
1844 static void __clear_buddies_last(struct sched_entity *se)
1846 for_each_sched_entity(se) {
1847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1848 if (cfs_rq->last == se)
1849 cfs_rq->last = NULL;
1855 static void __clear_buddies_next(struct sched_entity *se)
1857 for_each_sched_entity(se) {
1858 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1859 if (cfs_rq->next == se)
1860 cfs_rq->next = NULL;
1866 static void __clear_buddies_skip(struct sched_entity *se)
1868 for_each_sched_entity(se) {
1869 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1870 if (cfs_rq->skip == se)
1871 cfs_rq->skip = NULL;
1877 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1879 if (cfs_rq->last == se)
1880 __clear_buddies_last(se);
1882 if (cfs_rq->next == se)
1883 __clear_buddies_next(se);
1885 if (cfs_rq->skip == se)
1886 __clear_buddies_skip(se);
1889 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1892 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1895 * Update run-time statistics of the 'current'.
1897 update_curr(cfs_rq);
1898 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1900 update_stats_dequeue(cfs_rq, se);
1901 if (flags & DEQUEUE_SLEEP) {
1902 #ifdef CONFIG_SCHEDSTATS
1903 if (entity_is_task(se)) {
1904 struct task_struct *tsk = task_of(se);
1906 if (tsk->state & TASK_INTERRUPTIBLE)
1907 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1908 if (tsk->state & TASK_UNINTERRUPTIBLE)
1909 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1914 clear_buddies(cfs_rq, se);
1916 if (se != cfs_rq->curr)
1917 __dequeue_entity(cfs_rq, se);
1919 account_entity_dequeue(cfs_rq, se);
1922 * Normalize the entity after updating the min_vruntime because the
1923 * update can refer to the ->curr item and we need to reflect this
1924 * movement in our normalized position.
1926 if (!(flags & DEQUEUE_SLEEP))
1927 se->vruntime -= cfs_rq->min_vruntime;
1929 /* return excess runtime on last dequeue */
1930 return_cfs_rq_runtime(cfs_rq);
1932 update_min_vruntime(cfs_rq);
1933 update_cfs_shares(cfs_rq);
1937 * Preempt the current task with a newly woken task if needed:
1940 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1942 unsigned long ideal_runtime, delta_exec;
1943 struct sched_entity *se;
1946 ideal_runtime = sched_slice(cfs_rq, curr);
1947 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1948 if (delta_exec > ideal_runtime) {
1949 resched_task(rq_of(cfs_rq)->curr);
1951 * The current task ran long enough, ensure it doesn't get
1952 * re-elected due to buddy favours.
1954 clear_buddies(cfs_rq, curr);
1959 * Ensure that a task that missed wakeup preemption by a
1960 * narrow margin doesn't have to wait for a full slice.
1961 * This also mitigates buddy induced latencies under load.
1963 if (delta_exec < sysctl_sched_min_granularity)
1966 se = __pick_first_entity(cfs_rq);
1967 delta = curr->vruntime - se->vruntime;
1972 if (delta > ideal_runtime)
1973 resched_task(rq_of(cfs_rq)->curr);
1977 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1979 /* 'current' is not kept within the tree. */
1982 * Any task has to be enqueued before it get to execute on
1983 * a CPU. So account for the time it spent waiting on the
1986 update_stats_wait_end(cfs_rq, se);
1987 __dequeue_entity(cfs_rq, se);
1990 update_stats_curr_start(cfs_rq, se);
1992 #ifdef CONFIG_SCHEDSTATS
1994 * Track our maximum slice length, if the CPU's load is at
1995 * least twice that of our own weight (i.e. dont track it
1996 * when there are only lesser-weight tasks around):
1998 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1999 se->statistics.slice_max = max(se->statistics.slice_max,
2000 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2003 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2007 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2010 * Pick the next process, keeping these things in mind, in this order:
2011 * 1) keep things fair between processes/task groups
2012 * 2) pick the "next" process, since someone really wants that to run
2013 * 3) pick the "last" process, for cache locality
2014 * 4) do not run the "skip" process, if something else is available
2016 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2018 struct sched_entity *se = __pick_first_entity(cfs_rq);
2019 struct sched_entity *left = se;
2022 * Avoid running the skip buddy, if running something else can
2023 * be done without getting too unfair.
2025 if (cfs_rq->skip == se) {
2026 struct sched_entity *second = __pick_next_entity(se);
2027 if (second && wakeup_preempt_entity(second, left) < 1)
2032 * Prefer last buddy, try to return the CPU to a preempted task.
2034 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2038 * Someone really wants this to run. If it's not unfair, run it.
2040 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2043 clear_buddies(cfs_rq, se);
2048 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2050 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2053 * If still on the runqueue then deactivate_task()
2054 * was not called and update_curr() has to be done:
2057 update_curr(cfs_rq);
2059 /* throttle cfs_rqs exceeding runtime */
2060 check_cfs_rq_runtime(cfs_rq);
2062 check_spread(cfs_rq, prev);
2064 update_stats_wait_start(cfs_rq, prev);
2065 /* Put 'current' back into the tree. */
2066 __enqueue_entity(cfs_rq, prev);
2067 /* in !on_rq case, update occurred at dequeue */
2068 update_entity_load_avg(prev, 1);
2070 cfs_rq->curr = NULL;
2074 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2077 * Update run-time statistics of the 'current'.
2079 update_curr(cfs_rq);
2082 * Ensure that runnable average is periodically updated.
2084 update_entity_load_avg(curr, 1);
2085 update_cfs_rq_blocked_load(cfs_rq, 1);
2086 update_cfs_shares(cfs_rq);
2088 #ifdef CONFIG_SCHED_HRTICK
2090 * queued ticks are scheduled to match the slice, so don't bother
2091 * validating it and just reschedule.
2094 resched_task(rq_of(cfs_rq)->curr);
2098 * don't let the period tick interfere with the hrtick preemption
2100 if (!sched_feat(DOUBLE_TICK) &&
2101 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2105 if (cfs_rq->nr_running > 1)
2106 check_preempt_tick(cfs_rq, curr);
2110 /**************************************************
2111 * CFS bandwidth control machinery
2114 #ifdef CONFIG_CFS_BANDWIDTH
2116 #ifdef HAVE_JUMP_LABEL
2117 static struct static_key __cfs_bandwidth_used;
2119 static inline bool cfs_bandwidth_used(void)
2121 return static_key_false(&__cfs_bandwidth_used);
2124 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2126 /* only need to count groups transitioning between enabled/!enabled */
2127 if (enabled && !was_enabled)
2128 static_key_slow_inc(&__cfs_bandwidth_used);
2129 else if (!enabled && was_enabled)
2130 static_key_slow_dec(&__cfs_bandwidth_used);
2132 #else /* HAVE_JUMP_LABEL */
2133 static bool cfs_bandwidth_used(void)
2138 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2139 #endif /* HAVE_JUMP_LABEL */
2142 * default period for cfs group bandwidth.
2143 * default: 0.1s, units: nanoseconds
2145 static inline u64 default_cfs_period(void)
2147 return 100000000ULL;
2150 static inline u64 sched_cfs_bandwidth_slice(void)
2152 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2156 * Replenish runtime according to assigned quota and update expiration time.
2157 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2158 * additional synchronization around rq->lock.
2160 * requires cfs_b->lock
2162 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2166 if (cfs_b->quota == RUNTIME_INF)
2169 now = sched_clock_cpu(smp_processor_id());
2170 cfs_b->runtime = cfs_b->quota;
2171 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2174 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2176 return &tg->cfs_bandwidth;
2179 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2180 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2182 if (unlikely(cfs_rq->throttle_count))
2183 return cfs_rq->throttled_clock_task;
2185 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2188 /* returns 0 on failure to allocate runtime */
2189 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2191 struct task_group *tg = cfs_rq->tg;
2192 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2193 u64 amount = 0, min_amount, expires;
2195 /* note: this is a positive sum as runtime_remaining <= 0 */
2196 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2198 raw_spin_lock(&cfs_b->lock);
2199 if (cfs_b->quota == RUNTIME_INF)
2200 amount = min_amount;
2203 * If the bandwidth pool has become inactive, then at least one
2204 * period must have elapsed since the last consumption.
2205 * Refresh the global state and ensure bandwidth timer becomes
2208 if (!cfs_b->timer_active) {
2209 __refill_cfs_bandwidth_runtime(cfs_b);
2210 __start_cfs_bandwidth(cfs_b);
2213 if (cfs_b->runtime > 0) {
2214 amount = min(cfs_b->runtime, min_amount);
2215 cfs_b->runtime -= amount;
2219 expires = cfs_b->runtime_expires;
2220 raw_spin_unlock(&cfs_b->lock);
2222 cfs_rq->runtime_remaining += amount;
2224 * we may have advanced our local expiration to account for allowed
2225 * spread between our sched_clock and the one on which runtime was
2228 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2229 cfs_rq->runtime_expires = expires;
2231 return cfs_rq->runtime_remaining > 0;
2235 * Note: This depends on the synchronization provided by sched_clock and the
2236 * fact that rq->clock snapshots this value.
2238 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2240 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2242 /* if the deadline is ahead of our clock, nothing to do */
2243 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2246 if (cfs_rq->runtime_remaining < 0)
2250 * If the local deadline has passed we have to consider the
2251 * possibility that our sched_clock is 'fast' and the global deadline
2252 * has not truly expired.
2254 * Fortunately we can check determine whether this the case by checking
2255 * whether the global deadline has advanced.
2258 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2259 /* extend local deadline, drift is bounded above by 2 ticks */
2260 cfs_rq->runtime_expires += TICK_NSEC;
2262 /* global deadline is ahead, expiration has passed */
2263 cfs_rq->runtime_remaining = 0;
2267 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2268 unsigned long delta_exec)
2270 /* dock delta_exec before expiring quota (as it could span periods) */
2271 cfs_rq->runtime_remaining -= delta_exec;
2272 expire_cfs_rq_runtime(cfs_rq);
2274 if (likely(cfs_rq->runtime_remaining > 0))
2278 * if we're unable to extend our runtime we resched so that the active
2279 * hierarchy can be throttled
2281 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2282 resched_task(rq_of(cfs_rq)->curr);
2285 static __always_inline
2286 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2288 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2291 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2294 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2296 return cfs_bandwidth_used() && cfs_rq->throttled;
2299 /* check whether cfs_rq, or any parent, is throttled */
2300 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2302 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2306 * Ensure that neither of the group entities corresponding to src_cpu or
2307 * dest_cpu are members of a throttled hierarchy when performing group
2308 * load-balance operations.
2310 static inline int throttled_lb_pair(struct task_group *tg,
2311 int src_cpu, int dest_cpu)
2313 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2315 src_cfs_rq = tg->cfs_rq[src_cpu];
2316 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2318 return throttled_hierarchy(src_cfs_rq) ||
2319 throttled_hierarchy(dest_cfs_rq);
2322 /* updated child weight may affect parent so we have to do this bottom up */
2323 static int tg_unthrottle_up(struct task_group *tg, void *data)
2325 struct rq *rq = data;
2326 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2328 cfs_rq->throttle_count--;
2330 if (!cfs_rq->throttle_count) {
2331 /* adjust cfs_rq_clock_task() */
2332 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2333 cfs_rq->throttled_clock_task;
2340 static int tg_throttle_down(struct task_group *tg, void *data)
2342 struct rq *rq = data;
2343 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2345 /* group is entering throttled state, stop time */
2346 if (!cfs_rq->throttle_count)
2347 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2348 cfs_rq->throttle_count++;
2353 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2355 struct rq *rq = rq_of(cfs_rq);
2356 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2357 struct sched_entity *se;
2358 long task_delta, dequeue = 1;
2360 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2362 /* freeze hierarchy runnable averages while throttled */
2364 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2367 task_delta = cfs_rq->h_nr_running;
2368 for_each_sched_entity(se) {
2369 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2370 /* throttled entity or throttle-on-deactivate */
2375 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2376 qcfs_rq->h_nr_running -= task_delta;
2378 if (qcfs_rq->load.weight)
2383 rq->nr_running -= task_delta;
2385 cfs_rq->throttled = 1;
2386 cfs_rq->throttled_clock = rq_clock(rq);
2387 raw_spin_lock(&cfs_b->lock);
2388 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2389 raw_spin_unlock(&cfs_b->lock);
2392 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2394 struct rq *rq = rq_of(cfs_rq);
2395 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2396 struct sched_entity *se;
2400 se = cfs_rq->tg->se[cpu_of(rq)];
2402 cfs_rq->throttled = 0;
2404 update_rq_clock(rq);
2406 raw_spin_lock(&cfs_b->lock);
2407 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2408 list_del_rcu(&cfs_rq->throttled_list);
2409 raw_spin_unlock(&cfs_b->lock);
2411 /* update hierarchical throttle state */
2412 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2414 if (!cfs_rq->load.weight)
2417 task_delta = cfs_rq->h_nr_running;
2418 for_each_sched_entity(se) {
2422 cfs_rq = cfs_rq_of(se);
2424 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2425 cfs_rq->h_nr_running += task_delta;
2427 if (cfs_rq_throttled(cfs_rq))
2432 rq->nr_running += task_delta;
2434 /* determine whether we need to wake up potentially idle cpu */
2435 if (rq->curr == rq->idle && rq->cfs.nr_running)
2436 resched_task(rq->curr);
2439 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2440 u64 remaining, u64 expires)
2442 struct cfs_rq *cfs_rq;
2443 u64 runtime = remaining;
2446 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2448 struct rq *rq = rq_of(cfs_rq);
2450 raw_spin_lock(&rq->lock);
2451 if (!cfs_rq_throttled(cfs_rq))
2454 runtime = -cfs_rq->runtime_remaining + 1;
2455 if (runtime > remaining)
2456 runtime = remaining;
2457 remaining -= runtime;
2459 cfs_rq->runtime_remaining += runtime;
2460 cfs_rq->runtime_expires = expires;
2462 /* we check whether we're throttled above */
2463 if (cfs_rq->runtime_remaining > 0)
2464 unthrottle_cfs_rq(cfs_rq);
2467 raw_spin_unlock(&rq->lock);
2478 * Responsible for refilling a task_group's bandwidth and unthrottling its
2479 * cfs_rqs as appropriate. If there has been no activity within the last
2480 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2481 * used to track this state.
2483 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2485 u64 runtime, runtime_expires;
2486 int idle = 1, throttled;
2488 raw_spin_lock(&cfs_b->lock);
2489 /* no need to continue the timer with no bandwidth constraint */
2490 if (cfs_b->quota == RUNTIME_INF)
2493 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2494 /* idle depends on !throttled (for the case of a large deficit) */
2495 idle = cfs_b->idle && !throttled;
2496 cfs_b->nr_periods += overrun;
2498 /* if we're going inactive then everything else can be deferred */
2502 __refill_cfs_bandwidth_runtime(cfs_b);
2505 /* mark as potentially idle for the upcoming period */
2510 /* account preceding periods in which throttling occurred */
2511 cfs_b->nr_throttled += overrun;
2514 * There are throttled entities so we must first use the new bandwidth
2515 * to unthrottle them before making it generally available. This
2516 * ensures that all existing debts will be paid before a new cfs_rq is
2519 runtime = cfs_b->runtime;
2520 runtime_expires = cfs_b->runtime_expires;
2524 * This check is repeated as we are holding onto the new bandwidth
2525 * while we unthrottle. This can potentially race with an unthrottled
2526 * group trying to acquire new bandwidth from the global pool.
2528 while (throttled && runtime > 0) {
2529 raw_spin_unlock(&cfs_b->lock);
2530 /* we can't nest cfs_b->lock while distributing bandwidth */
2531 runtime = distribute_cfs_runtime(cfs_b, runtime,
2533 raw_spin_lock(&cfs_b->lock);
2535 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2538 /* return (any) remaining runtime */
2539 cfs_b->runtime = runtime;
2541 * While we are ensured activity in the period following an
2542 * unthrottle, this also covers the case in which the new bandwidth is
2543 * insufficient to cover the existing bandwidth deficit. (Forcing the
2544 * timer to remain active while there are any throttled entities.)
2549 cfs_b->timer_active = 0;
2550 raw_spin_unlock(&cfs_b->lock);
2555 /* a cfs_rq won't donate quota below this amount */
2556 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2557 /* minimum remaining period time to redistribute slack quota */
2558 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2559 /* how long we wait to gather additional slack before distributing */
2560 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2562 /* are we near the end of the current quota period? */
2563 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2565 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2568 /* if the call-back is running a quota refresh is already occurring */
2569 if (hrtimer_callback_running(refresh_timer))
2572 /* is a quota refresh about to occur? */
2573 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2574 if (remaining < min_expire)
2580 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2582 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2584 /* if there's a quota refresh soon don't bother with slack */
2585 if (runtime_refresh_within(cfs_b, min_left))
2588 start_bandwidth_timer(&cfs_b->slack_timer,
2589 ns_to_ktime(cfs_bandwidth_slack_period));
2592 /* we know any runtime found here is valid as update_curr() precedes return */
2593 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2595 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2596 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2598 if (slack_runtime <= 0)
2601 raw_spin_lock(&cfs_b->lock);
2602 if (cfs_b->quota != RUNTIME_INF &&
2603 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2604 cfs_b->runtime += slack_runtime;
2606 /* we are under rq->lock, defer unthrottling using a timer */
2607 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2608 !list_empty(&cfs_b->throttled_cfs_rq))
2609 start_cfs_slack_bandwidth(cfs_b);
2611 raw_spin_unlock(&cfs_b->lock);
2613 /* even if it's not valid for return we don't want to try again */
2614 cfs_rq->runtime_remaining -= slack_runtime;
2617 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2619 if (!cfs_bandwidth_used())
2622 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2625 __return_cfs_rq_runtime(cfs_rq);
2629 * This is done with a timer (instead of inline with bandwidth return) since
2630 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2632 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2634 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2637 /* confirm we're still not at a refresh boundary */
2638 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2641 raw_spin_lock(&cfs_b->lock);
2642 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2643 runtime = cfs_b->runtime;
2646 expires = cfs_b->runtime_expires;
2647 raw_spin_unlock(&cfs_b->lock);
2652 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2654 raw_spin_lock(&cfs_b->lock);
2655 if (expires == cfs_b->runtime_expires)
2656 cfs_b->runtime = runtime;
2657 raw_spin_unlock(&cfs_b->lock);
2661 * When a group wakes up we want to make sure that its quota is not already
2662 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2663 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2665 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2667 if (!cfs_bandwidth_used())
2670 /* an active group must be handled by the update_curr()->put() path */
2671 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2674 /* ensure the group is not already throttled */
2675 if (cfs_rq_throttled(cfs_rq))
2678 /* update runtime allocation */
2679 account_cfs_rq_runtime(cfs_rq, 0);
2680 if (cfs_rq->runtime_remaining <= 0)
2681 throttle_cfs_rq(cfs_rq);
2684 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2685 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2687 if (!cfs_bandwidth_used())
2690 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2694 * it's possible for a throttled entity to be forced into a running
2695 * state (e.g. set_curr_task), in this case we're finished.
2697 if (cfs_rq_throttled(cfs_rq))
2700 throttle_cfs_rq(cfs_rq);
2703 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2705 struct cfs_bandwidth *cfs_b =
2706 container_of(timer, struct cfs_bandwidth, slack_timer);
2707 do_sched_cfs_slack_timer(cfs_b);
2709 return HRTIMER_NORESTART;
2712 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2714 struct cfs_bandwidth *cfs_b =
2715 container_of(timer, struct cfs_bandwidth, period_timer);
2721 now = hrtimer_cb_get_time(timer);
2722 overrun = hrtimer_forward(timer, now, cfs_b->period);
2727 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2730 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2733 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2735 raw_spin_lock_init(&cfs_b->lock);
2737 cfs_b->quota = RUNTIME_INF;
2738 cfs_b->period = ns_to_ktime(default_cfs_period());
2740 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2741 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2742 cfs_b->period_timer.function = sched_cfs_period_timer;
2743 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2744 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2747 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2749 cfs_rq->runtime_enabled = 0;
2750 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2753 /* requires cfs_b->lock, may release to reprogram timer */
2754 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2757 * The timer may be active because we're trying to set a new bandwidth
2758 * period or because we're racing with the tear-down path
2759 * (timer_active==0 becomes visible before the hrtimer call-back
2760 * terminates). In either case we ensure that it's re-programmed
2762 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2763 raw_spin_unlock(&cfs_b->lock);
2764 /* ensure cfs_b->lock is available while we wait */
2765 hrtimer_cancel(&cfs_b->period_timer);
2767 raw_spin_lock(&cfs_b->lock);
2768 /* if someone else restarted the timer then we're done */
2769 if (cfs_b->timer_active)
2773 cfs_b->timer_active = 1;
2774 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2777 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2779 hrtimer_cancel(&cfs_b->period_timer);
2780 hrtimer_cancel(&cfs_b->slack_timer);
2783 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2785 struct cfs_rq *cfs_rq;
2787 for_each_leaf_cfs_rq(rq, cfs_rq) {
2788 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2790 if (!cfs_rq->runtime_enabled)
2794 * clock_task is not advancing so we just need to make sure
2795 * there's some valid quota amount
2797 cfs_rq->runtime_remaining = cfs_b->quota;
2798 if (cfs_rq_throttled(cfs_rq))
2799 unthrottle_cfs_rq(cfs_rq);
2803 #else /* CONFIG_CFS_BANDWIDTH */
2804 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2806 return rq_clock_task(rq_of(cfs_rq));
2809 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2810 unsigned long delta_exec) {}
2811 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2812 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2813 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2815 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2820 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2825 static inline int throttled_lb_pair(struct task_group *tg,
2826 int src_cpu, int dest_cpu)
2831 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2833 #ifdef CONFIG_FAIR_GROUP_SCHED
2834 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2837 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2841 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2842 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2844 #endif /* CONFIG_CFS_BANDWIDTH */
2846 /**************************************************
2847 * CFS operations on tasks:
2850 #ifdef CONFIG_SCHED_HRTICK
2851 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2853 struct sched_entity *se = &p->se;
2854 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2856 WARN_ON(task_rq(p) != rq);
2858 if (cfs_rq->nr_running > 1) {
2859 u64 slice = sched_slice(cfs_rq, se);
2860 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2861 s64 delta = slice - ran;
2870 * Don't schedule slices shorter than 10000ns, that just
2871 * doesn't make sense. Rely on vruntime for fairness.
2874 delta = max_t(s64, 10000LL, delta);
2876 hrtick_start(rq, delta);
2881 * called from enqueue/dequeue and updates the hrtick when the
2882 * current task is from our class and nr_running is low enough
2885 static void hrtick_update(struct rq *rq)
2887 struct task_struct *curr = rq->curr;
2889 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2892 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2893 hrtick_start_fair(rq, curr);
2895 #else /* !CONFIG_SCHED_HRTICK */
2897 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2901 static inline void hrtick_update(struct rq *rq)
2907 * The enqueue_task method is called before nr_running is
2908 * increased. Here we update the fair scheduling stats and
2909 * then put the task into the rbtree:
2912 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2914 struct cfs_rq *cfs_rq;
2915 struct sched_entity *se = &p->se;
2917 for_each_sched_entity(se) {
2920 cfs_rq = cfs_rq_of(se);
2921 enqueue_entity(cfs_rq, se, flags);
2924 * end evaluation on encountering a throttled cfs_rq
2926 * note: in the case of encountering a throttled cfs_rq we will
2927 * post the final h_nr_running increment below.
2929 if (cfs_rq_throttled(cfs_rq))
2931 cfs_rq->h_nr_running++;
2933 flags = ENQUEUE_WAKEUP;
2936 for_each_sched_entity(se) {
2937 cfs_rq = cfs_rq_of(se);
2938 cfs_rq->h_nr_running++;
2940 if (cfs_rq_throttled(cfs_rq))
2943 update_cfs_shares(cfs_rq);
2944 update_entity_load_avg(se, 1);
2948 update_rq_runnable_avg(rq, rq->nr_running);
2954 static void set_next_buddy(struct sched_entity *se);
2957 * The dequeue_task method is called before nr_running is
2958 * decreased. We remove the task from the rbtree and
2959 * update the fair scheduling stats:
2961 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2963 struct cfs_rq *cfs_rq;
2964 struct sched_entity *se = &p->se;
2965 int task_sleep = flags & DEQUEUE_SLEEP;
2967 for_each_sched_entity(se) {
2968 cfs_rq = cfs_rq_of(se);
2969 dequeue_entity(cfs_rq, se, flags);
2972 * end evaluation on encountering a throttled cfs_rq
2974 * note: in the case of encountering a throttled cfs_rq we will
2975 * post the final h_nr_running decrement below.
2977 if (cfs_rq_throttled(cfs_rq))
2979 cfs_rq->h_nr_running--;
2981 /* Don't dequeue parent if it has other entities besides us */
2982 if (cfs_rq->load.weight) {
2984 * Bias pick_next to pick a task from this cfs_rq, as
2985 * p is sleeping when it is within its sched_slice.
2987 if (task_sleep && parent_entity(se))
2988 set_next_buddy(parent_entity(se));
2990 /* avoid re-evaluating load for this entity */
2991 se = parent_entity(se);
2994 flags |= DEQUEUE_SLEEP;
2997 for_each_sched_entity(se) {
2998 cfs_rq = cfs_rq_of(se);
2999 cfs_rq->h_nr_running--;
3001 if (cfs_rq_throttled(cfs_rq))
3004 update_cfs_shares(cfs_rq);
3005 update_entity_load_avg(se, 1);
3010 update_rq_runnable_avg(rq, 1);
3016 /* Used instead of source_load when we know the type == 0 */
3017 static unsigned long weighted_cpuload(const int cpu)
3019 return cpu_rq(cpu)->cfs.runnable_load_avg;
3023 * Return a low guess at the load of a migration-source cpu weighted
3024 * according to the scheduling class and "nice" value.
3026 * We want to under-estimate the load of migration sources, to
3027 * balance conservatively.
3029 static unsigned long source_load(int cpu, int type)
3031 struct rq *rq = cpu_rq(cpu);
3032 unsigned long total = weighted_cpuload(cpu);
3034 if (type == 0 || !sched_feat(LB_BIAS))
3037 return min(rq->cpu_load[type-1], total);
3041 * Return a high guess at the load of a migration-target cpu weighted
3042 * according to the scheduling class and "nice" value.
3044 static unsigned long target_load(int cpu, int type)
3046 struct rq *rq = cpu_rq(cpu);
3047 unsigned long total = weighted_cpuload(cpu);
3049 if (type == 0 || !sched_feat(LB_BIAS))
3052 return max(rq->cpu_load[type-1], total);
3055 static unsigned long power_of(int cpu)
3057 return cpu_rq(cpu)->cpu_power;
3060 static unsigned long cpu_avg_load_per_task(int cpu)
3062 struct rq *rq = cpu_rq(cpu);
3063 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3064 unsigned long load_avg = rq->cfs.runnable_load_avg;
3067 return load_avg / nr_running;
3072 static void record_wakee(struct task_struct *p)
3075 * Rough decay (wiping) for cost saving, don't worry
3076 * about the boundary, really active task won't care
3079 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3080 current->wakee_flips = 0;
3081 current->wakee_flip_decay_ts = jiffies;
3084 if (current->last_wakee != p) {
3085 current->last_wakee = p;
3086 current->wakee_flips++;
3090 static void task_waking_fair(struct task_struct *p)
3092 struct sched_entity *se = &p->se;
3093 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3096 #ifndef CONFIG_64BIT
3097 u64 min_vruntime_copy;
3100 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3102 min_vruntime = cfs_rq->min_vruntime;
3103 } while (min_vruntime != min_vruntime_copy);
3105 min_vruntime = cfs_rq->min_vruntime;
3108 se->vruntime -= min_vruntime;
3112 #ifdef CONFIG_FAIR_GROUP_SCHED
3114 * effective_load() calculates the load change as seen from the root_task_group
3116 * Adding load to a group doesn't make a group heavier, but can cause movement
3117 * of group shares between cpus. Assuming the shares were perfectly aligned one
3118 * can calculate the shift in shares.
3120 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3121 * on this @cpu and results in a total addition (subtraction) of @wg to the
3122 * total group weight.
3124 * Given a runqueue weight distribution (rw_i) we can compute a shares
3125 * distribution (s_i) using:
3127 * s_i = rw_i / \Sum rw_j (1)
3129 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3130 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3131 * shares distribution (s_i):
3133 * rw_i = { 2, 4, 1, 0 }
3134 * s_i = { 2/7, 4/7, 1/7, 0 }
3136 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3137 * task used to run on and the CPU the waker is running on), we need to
3138 * compute the effect of waking a task on either CPU and, in case of a sync
3139 * wakeup, compute the effect of the current task going to sleep.
3141 * So for a change of @wl to the local @cpu with an overall group weight change
3142 * of @wl we can compute the new shares distribution (s'_i) using:
3144 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3146 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3147 * differences in waking a task to CPU 0. The additional task changes the
3148 * weight and shares distributions like:
3150 * rw'_i = { 3, 4, 1, 0 }
3151 * s'_i = { 3/8, 4/8, 1/8, 0 }
3153 * We can then compute the difference in effective weight by using:
3155 * dw_i = S * (s'_i - s_i) (3)
3157 * Where 'S' is the group weight as seen by its parent.
3159 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3160 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3161 * 4/7) times the weight of the group.
3163 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3165 struct sched_entity *se = tg->se[cpu];
3167 if (!tg->parent) /* the trivial, non-cgroup case */
3170 for_each_sched_entity(se) {
3176 * W = @wg + \Sum rw_j
3178 W = wg + calc_tg_weight(tg, se->my_q);
3183 w = se->my_q->load.weight + wl;
3186 * wl = S * s'_i; see (2)
3189 wl = (w * tg->shares) / W;
3194 * Per the above, wl is the new se->load.weight value; since
3195 * those are clipped to [MIN_SHARES, ...) do so now. See
3196 * calc_cfs_shares().
3198 if (wl < MIN_SHARES)
3202 * wl = dw_i = S * (s'_i - s_i); see (3)
3204 wl -= se->load.weight;
3207 * Recursively apply this logic to all parent groups to compute
3208 * the final effective load change on the root group. Since
3209 * only the @tg group gets extra weight, all parent groups can
3210 * only redistribute existing shares. @wl is the shift in shares
3211 * resulting from this level per the above.
3220 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3221 unsigned long wl, unsigned long wg)
3228 static int wake_wide(struct task_struct *p)
3230 int factor = this_cpu_read(sd_llc_size);
3233 * Yeah, it's the switching-frequency, could means many wakee or
3234 * rapidly switch, use factor here will just help to automatically
3235 * adjust the loose-degree, so bigger node will lead to more pull.
3237 if (p->wakee_flips > factor) {
3239 * wakee is somewhat hot, it needs certain amount of cpu
3240 * resource, so if waker is far more hot, prefer to leave
3243 if (current->wakee_flips > (factor * p->wakee_flips))
3250 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3252 s64 this_load, load;
3253 int idx, this_cpu, prev_cpu;
3254 unsigned long tl_per_task;
3255 struct task_group *tg;
3256 unsigned long weight;
3260 * If we wake multiple tasks be careful to not bounce
3261 * ourselves around too much.
3267 this_cpu = smp_processor_id();
3268 prev_cpu = task_cpu(p);
3269 load = source_load(prev_cpu, idx);
3270 this_load = target_load(this_cpu, idx);
3273 * If sync wakeup then subtract the (maximum possible)
3274 * effect of the currently running task from the load
3275 * of the current CPU:
3278 tg = task_group(current);
3279 weight = current->se.load.weight;
3281 this_load += effective_load(tg, this_cpu, -weight, -weight);
3282 load += effective_load(tg, prev_cpu, 0, -weight);
3286 weight = p->se.load.weight;
3289 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3290 * due to the sync cause above having dropped this_load to 0, we'll
3291 * always have an imbalance, but there's really nothing you can do
3292 * about that, so that's good too.
3294 * Otherwise check if either cpus are near enough in load to allow this
3295 * task to be woken on this_cpu.
3297 if (this_load > 0) {
3298 s64 this_eff_load, prev_eff_load;
3300 this_eff_load = 100;
3301 this_eff_load *= power_of(prev_cpu);
3302 this_eff_load *= this_load +
3303 effective_load(tg, this_cpu, weight, weight);
3305 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3306 prev_eff_load *= power_of(this_cpu);
3307 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3309 balanced = this_eff_load <= prev_eff_load;
3314 * If the currently running task will sleep within
3315 * a reasonable amount of time then attract this newly
3318 if (sync && balanced)
3321 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3322 tl_per_task = cpu_avg_load_per_task(this_cpu);
3325 (this_load <= load &&
3326 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3328 * This domain has SD_WAKE_AFFINE and
3329 * p is cache cold in this domain, and
3330 * there is no bad imbalance.
3332 schedstat_inc(sd, ttwu_move_affine);
3333 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3341 * find_idlest_group finds and returns the least busy CPU group within the
3344 static struct sched_group *
3345 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3346 int this_cpu, int load_idx)
3348 struct sched_group *idlest = NULL, *group = sd->groups;
3349 unsigned long min_load = ULONG_MAX, this_load = 0;
3350 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3353 unsigned long load, avg_load;
3357 /* Skip over this group if it has no CPUs allowed */
3358 if (!cpumask_intersects(sched_group_cpus(group),
3359 tsk_cpus_allowed(p)))
3362 local_group = cpumask_test_cpu(this_cpu,
3363 sched_group_cpus(group));
3365 /* Tally up the load of all CPUs in the group */
3368 for_each_cpu(i, sched_group_cpus(group)) {
3369 /* Bias balancing toward cpus of our domain */
3371 load = source_load(i, load_idx);
3373 load = target_load(i, load_idx);
3378 /* Adjust by relative CPU power of the group */
3379 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3382 this_load = avg_load;
3383 } else if (avg_load < min_load) {
3384 min_load = avg_load;
3387 } while (group = group->next, group != sd->groups);
3389 if (!idlest || 100*this_load < imbalance*min_load)
3395 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3398 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3400 unsigned long load, min_load = ULONG_MAX;
3404 /* Traverse only the allowed CPUs */
3405 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3406 load = weighted_cpuload(i);
3408 if (load < min_load || (load == min_load && i == this_cpu)) {
3418 * Try and locate an idle CPU in the sched_domain.
3420 static int select_idle_sibling(struct task_struct *p, int target)
3422 struct sched_domain *sd;
3423 struct sched_group *sg;
3424 int i = task_cpu(p);
3426 if (idle_cpu(target))
3430 * If the prevous cpu is cache affine and idle, don't be stupid.
3432 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3436 * Otherwise, iterate the domains and find an elegible idle cpu.
3438 sd = rcu_dereference(per_cpu(sd_llc, target));
3439 for_each_lower_domain(sd) {
3442 if (!cpumask_intersects(sched_group_cpus(sg),
3443 tsk_cpus_allowed(p)))
3446 for_each_cpu(i, sched_group_cpus(sg)) {
3447 if (i == target || !idle_cpu(i))
3451 target = cpumask_first_and(sched_group_cpus(sg),
3452 tsk_cpus_allowed(p));
3456 } while (sg != sd->groups);
3463 * sched_balance_self: balance the current task (running on cpu) in domains
3464 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3467 * Balance, ie. select the least loaded group.
3469 * Returns the target CPU number, or the same CPU if no balancing is needed.
3471 * preempt must be disabled.
3474 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3476 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3477 int cpu = smp_processor_id();
3478 int prev_cpu = task_cpu(p);
3480 int want_affine = 0;
3481 int sync = wake_flags & WF_SYNC;
3483 if (p->nr_cpus_allowed == 1)
3486 if (sd_flag & SD_BALANCE_WAKE) {
3487 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3493 for_each_domain(cpu, tmp) {
3494 if (!(tmp->flags & SD_LOAD_BALANCE))
3498 * If both cpu and prev_cpu are part of this domain,
3499 * cpu is a valid SD_WAKE_AFFINE target.
3501 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3502 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3507 if (tmp->flags & sd_flag)
3512 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3515 new_cpu = select_idle_sibling(p, prev_cpu);
3520 int load_idx = sd->forkexec_idx;
3521 struct sched_group *group;
3524 if (!(sd->flags & sd_flag)) {
3529 if (sd_flag & SD_BALANCE_WAKE)
3530 load_idx = sd->wake_idx;
3532 group = find_idlest_group(sd, p, cpu, load_idx);
3538 new_cpu = find_idlest_cpu(group, p, cpu);
3539 if (new_cpu == -1 || new_cpu == cpu) {
3540 /* Now try balancing at a lower domain level of cpu */
3545 /* Now try balancing at a lower domain level of new_cpu */
3547 weight = sd->span_weight;
3549 for_each_domain(cpu, tmp) {
3550 if (weight <= tmp->span_weight)
3552 if (tmp->flags & sd_flag)
3555 /* while loop will break here if sd == NULL */
3564 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3565 * cfs_rq_of(p) references at time of call are still valid and identify the
3566 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3567 * other assumptions, including the state of rq->lock, should be made.
3570 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3572 struct sched_entity *se = &p->se;
3573 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3576 * Load tracking: accumulate removed load so that it can be processed
3577 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3578 * to blocked load iff they have a positive decay-count. It can never
3579 * be negative here since on-rq tasks have decay-count == 0.
3581 if (se->avg.decay_count) {
3582 se->avg.decay_count = -__synchronize_entity_decay(se);
3583 atomic_long_add(se->avg.load_avg_contrib,
3584 &cfs_rq->removed_load);
3587 #endif /* CONFIG_SMP */
3589 static unsigned long
3590 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3592 unsigned long gran = sysctl_sched_wakeup_granularity;
3595 * Since its curr running now, convert the gran from real-time
3596 * to virtual-time in his units.
3598 * By using 'se' instead of 'curr' we penalize light tasks, so
3599 * they get preempted easier. That is, if 'se' < 'curr' then
3600 * the resulting gran will be larger, therefore penalizing the
3601 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3602 * be smaller, again penalizing the lighter task.
3604 * This is especially important for buddies when the leftmost
3605 * task is higher priority than the buddy.
3607 return calc_delta_fair(gran, se);
3611 * Should 'se' preempt 'curr'.
3625 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3627 s64 gran, vdiff = curr->vruntime - se->vruntime;
3632 gran = wakeup_gran(curr, se);
3639 static void set_last_buddy(struct sched_entity *se)
3641 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3644 for_each_sched_entity(se)
3645 cfs_rq_of(se)->last = se;
3648 static void set_next_buddy(struct sched_entity *se)
3650 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3653 for_each_sched_entity(se)
3654 cfs_rq_of(se)->next = se;
3657 static void set_skip_buddy(struct sched_entity *se)
3659 for_each_sched_entity(se)
3660 cfs_rq_of(se)->skip = se;
3664 * Preempt the current task with a newly woken task if needed:
3666 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3668 struct task_struct *curr = rq->curr;
3669 struct sched_entity *se = &curr->se, *pse = &p->se;
3670 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3671 int scale = cfs_rq->nr_running >= sched_nr_latency;
3672 int next_buddy_marked = 0;
3674 if (unlikely(se == pse))
3678 * This is possible from callers such as move_task(), in which we
3679 * unconditionally check_prempt_curr() after an enqueue (which may have
3680 * lead to a throttle). This both saves work and prevents false
3681 * next-buddy nomination below.
3683 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3686 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3687 set_next_buddy(pse);
3688 next_buddy_marked = 1;
3692 * We can come here with TIF_NEED_RESCHED already set from new task
3695 * Note: this also catches the edge-case of curr being in a throttled
3696 * group (e.g. via set_curr_task), since update_curr() (in the
3697 * enqueue of curr) will have resulted in resched being set. This
3698 * prevents us from potentially nominating it as a false LAST_BUDDY
3701 if (test_tsk_need_resched(curr))
3704 /* Idle tasks are by definition preempted by non-idle tasks. */
3705 if (unlikely(curr->policy == SCHED_IDLE) &&
3706 likely(p->policy != SCHED_IDLE))
3710 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3711 * is driven by the tick):
3713 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3716 find_matching_se(&se, &pse);
3717 update_curr(cfs_rq_of(se));
3719 if (wakeup_preempt_entity(se, pse) == 1) {
3721 * Bias pick_next to pick the sched entity that is
3722 * triggering this preemption.
3724 if (!next_buddy_marked)
3725 set_next_buddy(pse);
3734 * Only set the backward buddy when the current task is still
3735 * on the rq. This can happen when a wakeup gets interleaved
3736 * with schedule on the ->pre_schedule() or idle_balance()
3737 * point, either of which can * drop the rq lock.
3739 * Also, during early boot the idle thread is in the fair class,
3740 * for obvious reasons its a bad idea to schedule back to it.
3742 if (unlikely(!se->on_rq || curr == rq->idle))
3745 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3749 static struct task_struct *pick_next_task_fair(struct rq *rq)
3751 struct task_struct *p;
3752 struct cfs_rq *cfs_rq = &rq->cfs;
3753 struct sched_entity *se;
3755 if (!cfs_rq->nr_running)
3759 se = pick_next_entity(cfs_rq);
3760 set_next_entity(cfs_rq, se);
3761 cfs_rq = group_cfs_rq(se);
3765 if (hrtick_enabled(rq))
3766 hrtick_start_fair(rq, p);
3772 * Account for a descheduled task:
3774 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3776 struct sched_entity *se = &prev->se;
3777 struct cfs_rq *cfs_rq;
3779 for_each_sched_entity(se) {
3780 cfs_rq = cfs_rq_of(se);
3781 put_prev_entity(cfs_rq, se);
3786 * sched_yield() is very simple
3788 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3790 static void yield_task_fair(struct rq *rq)
3792 struct task_struct *curr = rq->curr;
3793 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3794 struct sched_entity *se = &curr->se;
3797 * Are we the only task in the tree?
3799 if (unlikely(rq->nr_running == 1))
3802 clear_buddies(cfs_rq, se);
3804 if (curr->policy != SCHED_BATCH) {
3805 update_rq_clock(rq);
3807 * Update run-time statistics of the 'current'.
3809 update_curr(cfs_rq);
3811 * Tell update_rq_clock() that we've just updated,
3812 * so we don't do microscopic update in schedule()
3813 * and double the fastpath cost.
3815 rq->skip_clock_update = 1;
3821 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3823 struct sched_entity *se = &p->se;
3825 /* throttled hierarchies are not runnable */
3826 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3829 /* Tell the scheduler that we'd really like pse to run next. */
3832 yield_task_fair(rq);
3838 /**************************************************
3839 * Fair scheduling class load-balancing methods.
3843 * The purpose of load-balancing is to achieve the same basic fairness the
3844 * per-cpu scheduler provides, namely provide a proportional amount of compute
3845 * time to each task. This is expressed in the following equation:
3847 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3849 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3850 * W_i,0 is defined as:
3852 * W_i,0 = \Sum_j w_i,j (2)
3854 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3855 * is derived from the nice value as per prio_to_weight[].
3857 * The weight average is an exponential decay average of the instantaneous
3860 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3862 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3863 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3864 * can also include other factors [XXX].
3866 * To achieve this balance we define a measure of imbalance which follows
3867 * directly from (1):
3869 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3871 * We them move tasks around to minimize the imbalance. In the continuous
3872 * function space it is obvious this converges, in the discrete case we get
3873 * a few fun cases generally called infeasible weight scenarios.
3876 * - infeasible weights;
3877 * - local vs global optima in the discrete case. ]
3882 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3883 * for all i,j solution, we create a tree of cpus that follows the hardware
3884 * topology where each level pairs two lower groups (or better). This results
3885 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3886 * tree to only the first of the previous level and we decrease the frequency
3887 * of load-balance at each level inv. proportional to the number of cpus in
3893 * \Sum { --- * --- * 2^i } = O(n) (5)
3895 * `- size of each group
3896 * | | `- number of cpus doing load-balance
3898 * `- sum over all levels
3900 * Coupled with a limit on how many tasks we can migrate every balance pass,
3901 * this makes (5) the runtime complexity of the balancer.
3903 * An important property here is that each CPU is still (indirectly) connected
3904 * to every other cpu in at most O(log n) steps:
3906 * The adjacency matrix of the resulting graph is given by:
3909 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3912 * And you'll find that:
3914 * A^(log_2 n)_i,j != 0 for all i,j (7)
3916 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3917 * The task movement gives a factor of O(m), giving a convergence complexity
3920 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3925 * In order to avoid CPUs going idle while there's still work to do, new idle
3926 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3927 * tree itself instead of relying on other CPUs to bring it work.
3929 * This adds some complexity to both (5) and (8) but it reduces the total idle
3937 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3940 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3945 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3947 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3949 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3952 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3953 * rewrite all of this once again.]
3956 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3958 #define LBF_ALL_PINNED 0x01
3959 #define LBF_NEED_BREAK 0x02
3960 #define LBF_DST_PINNED 0x04
3961 #define LBF_SOME_PINNED 0x08
3964 struct sched_domain *sd;
3972 struct cpumask *dst_grpmask;
3974 enum cpu_idle_type idle;
3976 /* The set of CPUs under consideration for load-balancing */
3977 struct cpumask *cpus;
3982 unsigned int loop_break;
3983 unsigned int loop_max;
3987 * move_task - move a task from one runqueue to another runqueue.
3988 * Both runqueues must be locked.
3990 static void move_task(struct task_struct *p, struct lb_env *env)
3992 deactivate_task(env->src_rq, p, 0);
3993 set_task_cpu(p, env->dst_cpu);
3994 activate_task(env->dst_rq, p, 0);
3995 check_preempt_curr(env->dst_rq, p, 0);
3999 * Is this task likely cache-hot:
4002 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4006 if (p->sched_class != &fair_sched_class)
4009 if (unlikely(p->policy == SCHED_IDLE))
4013 * Buddy candidates are cache hot:
4015 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4016 (&p->se == cfs_rq_of(&p->se)->next ||
4017 &p->se == cfs_rq_of(&p->se)->last))
4020 if (sysctl_sched_migration_cost == -1)
4022 if (sysctl_sched_migration_cost == 0)
4025 delta = now - p->se.exec_start;
4027 return delta < (s64)sysctl_sched_migration_cost;
4031 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4034 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4036 int tsk_cache_hot = 0;
4038 * We do not migrate tasks that are:
4039 * 1) throttled_lb_pair, or
4040 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4041 * 3) running (obviously), or
4042 * 4) are cache-hot on their current CPU.
4044 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4047 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4050 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4052 env->flags |= LBF_SOME_PINNED;
4055 * Remember if this task can be migrated to any other cpu in
4056 * our sched_group. We may want to revisit it if we couldn't
4057 * meet load balance goals by pulling other tasks on src_cpu.
4059 * Also avoid computing new_dst_cpu if we have already computed
4060 * one in current iteration.
4062 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4065 /* Prevent to re-select dst_cpu via env's cpus */
4066 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4067 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4068 env->flags |= LBF_DST_PINNED;
4069 env->new_dst_cpu = cpu;
4077 /* Record that we found atleast one task that could run on dst_cpu */
4078 env->flags &= ~LBF_ALL_PINNED;
4080 if (task_running(env->src_rq, p)) {
4081 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4086 * Aggressive migration if:
4087 * 1) task is cache cold, or
4088 * 2) too many balance attempts have failed.
4091 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4092 if (!tsk_cache_hot ||
4093 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4095 if (tsk_cache_hot) {
4096 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4097 schedstat_inc(p, se.statistics.nr_forced_migrations);
4103 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4108 * move_one_task tries to move exactly one task from busiest to this_rq, as
4109 * part of active balancing operations within "domain".
4110 * Returns 1 if successful and 0 otherwise.
4112 * Called with both runqueues locked.
4114 static int move_one_task(struct lb_env *env)
4116 struct task_struct *p, *n;
4118 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4119 if (!can_migrate_task(p, env))
4124 * Right now, this is only the second place move_task()
4125 * is called, so we can safely collect move_task()
4126 * stats here rather than inside move_task().
4128 schedstat_inc(env->sd, lb_gained[env->idle]);
4134 static unsigned long task_h_load(struct task_struct *p);
4136 static const unsigned int sched_nr_migrate_break = 32;
4139 * move_tasks tries to move up to imbalance weighted load from busiest to
4140 * this_rq, as part of a balancing operation within domain "sd".
4141 * Returns 1 if successful and 0 otherwise.
4143 * Called with both runqueues locked.
4145 static int move_tasks(struct lb_env *env)
4147 struct list_head *tasks = &env->src_rq->cfs_tasks;
4148 struct task_struct *p;
4152 if (env->imbalance <= 0)
4155 while (!list_empty(tasks)) {
4156 p = list_first_entry(tasks, struct task_struct, se.group_node);
4159 /* We've more or less seen every task there is, call it quits */
4160 if (env->loop > env->loop_max)
4163 /* take a breather every nr_migrate tasks */
4164 if (env->loop > env->loop_break) {
4165 env->loop_break += sched_nr_migrate_break;
4166 env->flags |= LBF_NEED_BREAK;
4170 if (!can_migrate_task(p, env))
4173 load = task_h_load(p);
4175 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4178 if ((load / 2) > env->imbalance)
4183 env->imbalance -= load;
4185 #ifdef CONFIG_PREEMPT
4187 * NEWIDLE balancing is a source of latency, so preemptible
4188 * kernels will stop after the first task is pulled to minimize
4189 * the critical section.
4191 if (env->idle == CPU_NEWLY_IDLE)
4196 * We only want to steal up to the prescribed amount of
4199 if (env->imbalance <= 0)
4204 list_move_tail(&p->se.group_node, tasks);
4208 * Right now, this is one of only two places move_task() is called,
4209 * so we can safely collect move_task() stats here rather than
4210 * inside move_task().
4212 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4217 #ifdef CONFIG_FAIR_GROUP_SCHED
4219 * update tg->load_weight by folding this cpu's load_avg
4221 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4223 struct sched_entity *se = tg->se[cpu];
4224 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4226 /* throttled entities do not contribute to load */
4227 if (throttled_hierarchy(cfs_rq))
4230 update_cfs_rq_blocked_load(cfs_rq, 1);
4233 update_entity_load_avg(se, 1);
4235 * We pivot on our runnable average having decayed to zero for
4236 * list removal. This generally implies that all our children
4237 * have also been removed (modulo rounding error or bandwidth
4238 * control); however, such cases are rare and we can fix these
4241 * TODO: fix up out-of-order children on enqueue.
4243 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4244 list_del_leaf_cfs_rq(cfs_rq);
4246 struct rq *rq = rq_of(cfs_rq);
4247 update_rq_runnable_avg(rq, rq->nr_running);
4251 static void update_blocked_averages(int cpu)
4253 struct rq *rq = cpu_rq(cpu);
4254 struct cfs_rq *cfs_rq;
4255 unsigned long flags;
4257 raw_spin_lock_irqsave(&rq->lock, flags);
4258 update_rq_clock(rq);
4260 * Iterates the task_group tree in a bottom up fashion, see
4261 * list_add_leaf_cfs_rq() for details.
4263 for_each_leaf_cfs_rq(rq, cfs_rq) {
4265 * Note: We may want to consider periodically releasing
4266 * rq->lock about these updates so that creating many task
4267 * groups does not result in continually extending hold time.
4269 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4272 raw_spin_unlock_irqrestore(&rq->lock, flags);
4276 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4277 * This needs to be done in a top-down fashion because the load of a child
4278 * group is a fraction of its parents load.
4280 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4282 struct rq *rq = rq_of(cfs_rq);
4283 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4284 unsigned long now = jiffies;
4287 if (cfs_rq->last_h_load_update == now)
4290 cfs_rq->h_load_next = NULL;
4291 for_each_sched_entity(se) {
4292 cfs_rq = cfs_rq_of(se);
4293 cfs_rq->h_load_next = se;
4294 if (cfs_rq->last_h_load_update == now)
4299 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4300 cfs_rq->last_h_load_update = now;
4303 while ((se = cfs_rq->h_load_next) != NULL) {
4304 load = cfs_rq->h_load;
4305 load = div64_ul(load * se->avg.load_avg_contrib,
4306 cfs_rq->runnable_load_avg + 1);
4307 cfs_rq = group_cfs_rq(se);
4308 cfs_rq->h_load = load;
4309 cfs_rq->last_h_load_update = now;
4313 static unsigned long task_h_load(struct task_struct *p)
4315 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4317 update_cfs_rq_h_load(cfs_rq);
4318 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4319 cfs_rq->runnable_load_avg + 1);
4322 static inline void update_blocked_averages(int cpu)
4326 static unsigned long task_h_load(struct task_struct *p)
4328 return p->se.avg.load_avg_contrib;
4332 /********** Helpers for find_busiest_group ************************/
4334 * sg_lb_stats - stats of a sched_group required for load_balancing
4336 struct sg_lb_stats {
4337 unsigned long avg_load; /*Avg load across the CPUs of the group */
4338 unsigned long group_load; /* Total load over the CPUs of the group */
4339 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4340 unsigned long load_per_task;
4341 unsigned long group_power;
4342 unsigned int sum_nr_running; /* Nr tasks running in the group */
4343 unsigned int group_capacity;
4344 unsigned int idle_cpus;
4345 unsigned int group_weight;
4346 int group_imb; /* Is there an imbalance in the group ? */
4347 int group_has_capacity; /* Is there extra capacity in the group? */
4351 * sd_lb_stats - Structure to store the statistics of a sched_domain
4352 * during load balancing.
4354 struct sd_lb_stats {
4355 struct sched_group *busiest; /* Busiest group in this sd */
4356 struct sched_group *local; /* Local group in this sd */
4357 unsigned long total_load; /* Total load of all groups in sd */
4358 unsigned long total_pwr; /* Total power of all groups in sd */
4359 unsigned long avg_load; /* Average load across all groups in sd */
4361 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4362 struct sg_lb_stats local_stat; /* Statistics of the local group */
4365 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4368 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4369 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4370 * We must however clear busiest_stat::avg_load because
4371 * update_sd_pick_busiest() reads this before assignment.
4373 *sds = (struct sd_lb_stats){
4385 * get_sd_load_idx - Obtain the load index for a given sched domain.
4386 * @sd: The sched_domain whose load_idx is to be obtained.
4387 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4389 * Return: The load index.
4391 static inline int get_sd_load_idx(struct sched_domain *sd,
4392 enum cpu_idle_type idle)
4398 load_idx = sd->busy_idx;
4401 case CPU_NEWLY_IDLE:
4402 load_idx = sd->newidle_idx;
4405 load_idx = sd->idle_idx;
4412 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4414 return SCHED_POWER_SCALE;
4417 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4419 return default_scale_freq_power(sd, cpu);
4422 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4424 unsigned long weight = sd->span_weight;
4425 unsigned long smt_gain = sd->smt_gain;
4432 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4434 return default_scale_smt_power(sd, cpu);
4437 static unsigned long scale_rt_power(int cpu)
4439 struct rq *rq = cpu_rq(cpu);
4440 u64 total, available, age_stamp, avg;
4443 * Since we're reading these variables without serialization make sure
4444 * we read them once before doing sanity checks on them.
4446 age_stamp = ACCESS_ONCE(rq->age_stamp);
4447 avg = ACCESS_ONCE(rq->rt_avg);
4449 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4451 if (unlikely(total < avg)) {
4452 /* Ensures that power won't end up being negative */
4455 available = total - avg;
4458 if (unlikely((s64)total < SCHED_POWER_SCALE))
4459 total = SCHED_POWER_SCALE;
4461 total >>= SCHED_POWER_SHIFT;
4463 return div_u64(available, total);
4466 static void update_cpu_power(struct sched_domain *sd, int cpu)
4468 unsigned long weight = sd->span_weight;
4469 unsigned long power = SCHED_POWER_SCALE;
4470 struct sched_group *sdg = sd->groups;
4472 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4473 if (sched_feat(ARCH_POWER))
4474 power *= arch_scale_smt_power(sd, cpu);
4476 power *= default_scale_smt_power(sd, cpu);
4478 power >>= SCHED_POWER_SHIFT;
4481 sdg->sgp->power_orig = power;
4483 if (sched_feat(ARCH_POWER))
4484 power *= arch_scale_freq_power(sd, cpu);
4486 power *= default_scale_freq_power(sd, cpu);
4488 power >>= SCHED_POWER_SHIFT;
4490 power *= scale_rt_power(cpu);
4491 power >>= SCHED_POWER_SHIFT;
4496 cpu_rq(cpu)->cpu_power = power;
4497 sdg->sgp->power = power;
4500 void update_group_power(struct sched_domain *sd, int cpu)
4502 struct sched_domain *child = sd->child;
4503 struct sched_group *group, *sdg = sd->groups;
4504 unsigned long power, power_orig;
4505 unsigned long interval;
4507 interval = msecs_to_jiffies(sd->balance_interval);
4508 interval = clamp(interval, 1UL, max_load_balance_interval);
4509 sdg->sgp->next_update = jiffies + interval;
4512 update_cpu_power(sd, cpu);
4516 power_orig = power = 0;
4518 if (child->flags & SD_OVERLAP) {
4520 * SD_OVERLAP domains cannot assume that child groups
4521 * span the current group.
4524 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4525 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4527 power_orig += sg->sgp->power_orig;
4528 power += sg->sgp->power;
4532 * !SD_OVERLAP domains can assume that child groups
4533 * span the current group.
4536 group = child->groups;
4538 power_orig += group->sgp->power_orig;
4539 power += group->sgp->power;
4540 group = group->next;
4541 } while (group != child->groups);
4544 sdg->sgp->power_orig = power_orig;
4545 sdg->sgp->power = power;
4549 * Try and fix up capacity for tiny siblings, this is needed when
4550 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4551 * which on its own isn't powerful enough.
4553 * See update_sd_pick_busiest() and check_asym_packing().
4556 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4559 * Only siblings can have significantly less than SCHED_POWER_SCALE
4561 if (!(sd->flags & SD_SHARE_CPUPOWER))
4565 * If ~90% of the cpu_power is still there, we're good.
4567 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4574 * Group imbalance indicates (and tries to solve) the problem where balancing
4575 * groups is inadequate due to tsk_cpus_allowed() constraints.
4577 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4578 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4581 * { 0 1 2 3 } { 4 5 6 7 }
4584 * If we were to balance group-wise we'd place two tasks in the first group and
4585 * two tasks in the second group. Clearly this is undesired as it will overload
4586 * cpu 3 and leave one of the cpus in the second group unused.
4588 * The current solution to this issue is detecting the skew in the first group
4589 * by noticing the lower domain failed to reach balance and had difficulty
4590 * moving tasks due to affinity constraints.
4592 * When this is so detected; this group becomes a candidate for busiest; see
4593 * update_sd_pick_busiest(). And calculcate_imbalance() and
4594 * find_busiest_group() avoid some of the usual balance conditions to allow it
4595 * to create an effective group imbalance.
4597 * This is a somewhat tricky proposition since the next run might not find the
4598 * group imbalance and decide the groups need to be balanced again. A most
4599 * subtle and fragile situation.
4602 static inline int sg_imbalanced(struct sched_group *group)
4604 return group->sgp->imbalance;
4608 * Compute the group capacity.
4610 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4611 * first dividing out the smt factor and computing the actual number of cores
4612 * and limit power unit capacity with that.
4614 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4616 unsigned int capacity, smt, cpus;
4617 unsigned int power, power_orig;
4619 power = group->sgp->power;
4620 power_orig = group->sgp->power_orig;
4621 cpus = group->group_weight;
4623 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4624 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4625 capacity = cpus / smt; /* cores */
4627 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4629 capacity = fix_small_capacity(env->sd, group);
4635 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4636 * @env: The load balancing environment.
4637 * @group: sched_group whose statistics are to be updated.
4638 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4639 * @local_group: Does group contain this_cpu.
4640 * @sgs: variable to hold the statistics for this group.
4642 static inline void update_sg_lb_stats(struct lb_env *env,
4643 struct sched_group *group, int load_idx,
4644 int local_group, struct sg_lb_stats *sgs)
4646 unsigned long nr_running;
4650 memset(sgs, 0, sizeof(*sgs));
4652 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4653 struct rq *rq = cpu_rq(i);
4655 nr_running = rq->nr_running;
4657 /* Bias balancing toward cpus of our domain */
4659 load = target_load(i, load_idx);
4661 load = source_load(i, load_idx);
4663 sgs->group_load += load;
4664 sgs->sum_nr_running += nr_running;
4665 sgs->sum_weighted_load += weighted_cpuload(i);
4670 /* Adjust by relative CPU power of the group */
4671 sgs->group_power = group->sgp->power;
4672 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4674 if (sgs->sum_nr_running)
4675 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4677 sgs->group_weight = group->group_weight;
4679 sgs->group_imb = sg_imbalanced(group);
4680 sgs->group_capacity = sg_capacity(env, group);
4682 if (sgs->group_capacity > sgs->sum_nr_running)
4683 sgs->group_has_capacity = 1;
4687 * update_sd_pick_busiest - return 1 on busiest group
4688 * @env: The load balancing environment.
4689 * @sds: sched_domain statistics
4690 * @sg: sched_group candidate to be checked for being the busiest
4691 * @sgs: sched_group statistics
4693 * Determine if @sg is a busier group than the previously selected
4696 * Return: %true if @sg is a busier group than the previously selected
4697 * busiest group. %false otherwise.
4699 static bool update_sd_pick_busiest(struct lb_env *env,
4700 struct sd_lb_stats *sds,
4701 struct sched_group *sg,
4702 struct sg_lb_stats *sgs)
4704 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4707 if (sgs->sum_nr_running > sgs->group_capacity)
4714 * ASYM_PACKING needs to move all the work to the lowest
4715 * numbered CPUs in the group, therefore mark all groups
4716 * higher than ourself as busy.
4718 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4719 env->dst_cpu < group_first_cpu(sg)) {
4723 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4731 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4732 * @env: The load balancing environment.
4733 * @balance: Should we balance.
4734 * @sds: variable to hold the statistics for this sched_domain.
4736 static inline void update_sd_lb_stats(struct lb_env *env,
4737 struct sd_lb_stats *sds)
4739 struct sched_domain *child = env->sd->child;
4740 struct sched_group *sg = env->sd->groups;
4741 struct sg_lb_stats tmp_sgs;
4742 int load_idx, prefer_sibling = 0;
4744 if (child && child->flags & SD_PREFER_SIBLING)
4747 load_idx = get_sd_load_idx(env->sd, env->idle);
4750 struct sg_lb_stats *sgs = &tmp_sgs;
4753 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4756 sgs = &sds->local_stat;
4758 if (env->idle != CPU_NEWLY_IDLE ||
4759 time_after_eq(jiffies, sg->sgp->next_update))
4760 update_group_power(env->sd, env->dst_cpu);
4763 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4769 * In case the child domain prefers tasks go to siblings
4770 * first, lower the sg capacity to one so that we'll try
4771 * and move all the excess tasks away. We lower the capacity
4772 * of a group only if the local group has the capacity to fit
4773 * these excess tasks, i.e. nr_running < group_capacity. The
4774 * extra check prevents the case where you always pull from the
4775 * heaviest group when it is already under-utilized (possible
4776 * with a large weight task outweighs the tasks on the system).
4778 if (prefer_sibling && sds->local &&
4779 sds->local_stat.group_has_capacity)
4780 sgs->group_capacity = min(sgs->group_capacity, 1U);
4782 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4784 sds->busiest_stat = *sgs;
4788 /* Now, start updating sd_lb_stats */
4789 sds->total_load += sgs->group_load;
4790 sds->total_pwr += sgs->group_power;
4793 } while (sg != env->sd->groups);
4797 * check_asym_packing - Check to see if the group is packed into the
4800 * This is primarily intended to used at the sibling level. Some
4801 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4802 * case of POWER7, it can move to lower SMT modes only when higher
4803 * threads are idle. When in lower SMT modes, the threads will
4804 * perform better since they share less core resources. Hence when we
4805 * have idle threads, we want them to be the higher ones.
4807 * This packing function is run on idle threads. It checks to see if
4808 * the busiest CPU in this domain (core in the P7 case) has a higher
4809 * CPU number than the packing function is being run on. Here we are
4810 * assuming lower CPU number will be equivalent to lower a SMT thread
4813 * Return: 1 when packing is required and a task should be moved to
4814 * this CPU. The amount of the imbalance is returned in *imbalance.
4816 * @env: The load balancing environment.
4817 * @sds: Statistics of the sched_domain which is to be packed
4819 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4823 if (!(env->sd->flags & SD_ASYM_PACKING))
4829 busiest_cpu = group_first_cpu(sds->busiest);
4830 if (env->dst_cpu > busiest_cpu)
4833 env->imbalance = DIV_ROUND_CLOSEST(
4834 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4841 * fix_small_imbalance - Calculate the minor imbalance that exists
4842 * amongst the groups of a sched_domain, during
4844 * @env: The load balancing environment.
4845 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4848 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4850 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4851 unsigned int imbn = 2;
4852 unsigned long scaled_busy_load_per_task;
4853 struct sg_lb_stats *local, *busiest;
4855 local = &sds->local_stat;
4856 busiest = &sds->busiest_stat;
4858 if (!local->sum_nr_running)
4859 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4860 else if (busiest->load_per_task > local->load_per_task)
4863 scaled_busy_load_per_task =
4864 (busiest->load_per_task * SCHED_POWER_SCALE) /
4865 busiest->group_power;
4867 if (busiest->avg_load + scaled_busy_load_per_task >=
4868 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4869 env->imbalance = busiest->load_per_task;
4874 * OK, we don't have enough imbalance to justify moving tasks,
4875 * however we may be able to increase total CPU power used by
4879 pwr_now += busiest->group_power *
4880 min(busiest->load_per_task, busiest->avg_load);
4881 pwr_now += local->group_power *
4882 min(local->load_per_task, local->avg_load);
4883 pwr_now /= SCHED_POWER_SCALE;
4885 /* Amount of load we'd subtract */
4886 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4887 busiest->group_power;
4888 if (busiest->avg_load > tmp) {
4889 pwr_move += busiest->group_power *
4890 min(busiest->load_per_task,
4891 busiest->avg_load - tmp);
4894 /* Amount of load we'd add */
4895 if (busiest->avg_load * busiest->group_power <
4896 busiest->load_per_task * SCHED_POWER_SCALE) {
4897 tmp = (busiest->avg_load * busiest->group_power) /
4900 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4903 pwr_move += local->group_power *
4904 min(local->load_per_task, local->avg_load + tmp);
4905 pwr_move /= SCHED_POWER_SCALE;
4907 /* Move if we gain throughput */
4908 if (pwr_move > pwr_now)
4909 env->imbalance = busiest->load_per_task;
4913 * calculate_imbalance - Calculate the amount of imbalance present within the
4914 * groups of a given sched_domain during load balance.
4915 * @env: load balance environment
4916 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4918 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4920 unsigned long max_pull, load_above_capacity = ~0UL;
4921 struct sg_lb_stats *local, *busiest;
4923 local = &sds->local_stat;
4924 busiest = &sds->busiest_stat;
4926 if (busiest->group_imb) {
4928 * In the group_imb case we cannot rely on group-wide averages
4929 * to ensure cpu-load equilibrium, look at wider averages. XXX
4931 busiest->load_per_task =
4932 min(busiest->load_per_task, sds->avg_load);
4936 * In the presence of smp nice balancing, certain scenarios can have
4937 * max load less than avg load(as we skip the groups at or below
4938 * its cpu_power, while calculating max_load..)
4940 if (busiest->avg_load <= sds->avg_load ||
4941 local->avg_load >= sds->avg_load) {
4943 return fix_small_imbalance(env, sds);
4946 if (!busiest->group_imb) {
4948 * Don't want to pull so many tasks that a group would go idle.
4949 * Except of course for the group_imb case, since then we might
4950 * have to drop below capacity to reach cpu-load equilibrium.
4952 load_above_capacity =
4953 (busiest->sum_nr_running - busiest->group_capacity);
4955 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4956 load_above_capacity /= busiest->group_power;
4960 * We're trying to get all the cpus to the average_load, so we don't
4961 * want to push ourselves above the average load, nor do we wish to
4962 * reduce the max loaded cpu below the average load. At the same time,
4963 * we also don't want to reduce the group load below the group capacity
4964 * (so that we can implement power-savings policies etc). Thus we look
4965 * for the minimum possible imbalance.
4967 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
4969 /* How much load to actually move to equalise the imbalance */
4970 env->imbalance = min(
4971 max_pull * busiest->group_power,
4972 (sds->avg_load - local->avg_load) * local->group_power
4973 ) / SCHED_POWER_SCALE;
4976 * if *imbalance is less than the average load per runnable task
4977 * there is no guarantee that any tasks will be moved so we'll have
4978 * a think about bumping its value to force at least one task to be
4981 if (env->imbalance < busiest->load_per_task)
4982 return fix_small_imbalance(env, sds);
4985 /******* find_busiest_group() helpers end here *********************/
4988 * find_busiest_group - Returns the busiest group within the sched_domain
4989 * if there is an imbalance. If there isn't an imbalance, and
4990 * the user has opted for power-savings, it returns a group whose
4991 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4992 * such a group exists.
4994 * Also calculates the amount of weighted load which should be moved
4995 * to restore balance.
4997 * @env: The load balancing environment.
4999 * Return: - The busiest group if imbalance exists.
5000 * - If no imbalance and user has opted for power-savings balance,
5001 * return the least loaded group whose CPUs can be
5002 * put to idle by rebalancing its tasks onto our group.
5004 static struct sched_group *find_busiest_group(struct lb_env *env)
5006 struct sg_lb_stats *local, *busiest;
5007 struct sd_lb_stats sds;
5009 init_sd_lb_stats(&sds);
5012 * Compute the various statistics relavent for load balancing at
5015 update_sd_lb_stats(env, &sds);
5016 local = &sds.local_stat;
5017 busiest = &sds.busiest_stat;
5019 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5020 check_asym_packing(env, &sds))
5023 /* There is no busy sibling group to pull tasks from */
5024 if (!sds.busiest || busiest->sum_nr_running == 0)
5027 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5030 * If the busiest group is imbalanced the below checks don't
5031 * work because they assume all things are equal, which typically
5032 * isn't true due to cpus_allowed constraints and the like.
5034 if (busiest->group_imb)
5037 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5038 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5039 !busiest->group_has_capacity)
5043 * If the local group is more busy than the selected busiest group
5044 * don't try and pull any tasks.
5046 if (local->avg_load >= busiest->avg_load)
5050 * Don't pull any tasks if this group is already above the domain
5053 if (local->avg_load >= sds.avg_load)
5056 if (env->idle == CPU_IDLE) {
5058 * This cpu is idle. If the busiest group load doesn't
5059 * have more tasks than the number of available cpu's and
5060 * there is no imbalance between this and busiest group
5061 * wrt to idle cpu's, it is balanced.
5063 if ((local->idle_cpus < busiest->idle_cpus) &&
5064 busiest->sum_nr_running <= busiest->group_weight)
5068 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5069 * imbalance_pct to be conservative.
5071 if (100 * busiest->avg_load <=
5072 env->sd->imbalance_pct * local->avg_load)
5077 /* Looks like there is an imbalance. Compute it */
5078 calculate_imbalance(env, &sds);
5087 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5089 static struct rq *find_busiest_queue(struct lb_env *env,
5090 struct sched_group *group)
5092 struct rq *busiest = NULL, *rq;
5093 unsigned long busiest_load = 0, busiest_power = 1;
5096 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5097 unsigned long power = power_of(i);
5098 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5103 capacity = fix_small_capacity(env->sd, group);
5106 wl = weighted_cpuload(i);
5109 * When comparing with imbalance, use weighted_cpuload()
5110 * which is not scaled with the cpu power.
5112 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5116 * For the load comparisons with the other cpu's, consider
5117 * the weighted_cpuload() scaled with the cpu power, so that
5118 * the load can be moved away from the cpu that is potentially
5119 * running at a lower capacity.
5121 * Thus we're looking for max(wl_i / power_i), crosswise
5122 * multiplication to rid ourselves of the division works out
5123 * to: wl_i * power_j > wl_j * power_i; where j is our
5126 if (wl * busiest_power > busiest_load * power) {
5128 busiest_power = power;
5137 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5138 * so long as it is large enough.
5140 #define MAX_PINNED_INTERVAL 512
5142 /* Working cpumask for load_balance and load_balance_newidle. */
5143 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5145 static int need_active_balance(struct lb_env *env)
5147 struct sched_domain *sd = env->sd;
5149 if (env->idle == CPU_NEWLY_IDLE) {
5152 * ASYM_PACKING needs to force migrate tasks from busy but
5153 * higher numbered CPUs in order to pack all tasks in the
5154 * lowest numbered CPUs.
5156 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5160 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5163 static int active_load_balance_cpu_stop(void *data);
5165 static int should_we_balance(struct lb_env *env)
5167 struct sched_group *sg = env->sd->groups;
5168 struct cpumask *sg_cpus, *sg_mask;
5169 int cpu, balance_cpu = -1;
5172 * In the newly idle case, we will allow all the cpu's
5173 * to do the newly idle load balance.
5175 if (env->idle == CPU_NEWLY_IDLE)
5178 sg_cpus = sched_group_cpus(sg);
5179 sg_mask = sched_group_mask(sg);
5180 /* Try to find first idle cpu */
5181 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5182 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5189 if (balance_cpu == -1)
5190 balance_cpu = group_balance_cpu(sg);
5193 * First idle cpu or the first cpu(busiest) in this sched group
5194 * is eligible for doing load balancing at this and above domains.
5196 return balance_cpu == env->dst_cpu;
5200 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5201 * tasks if there is an imbalance.
5203 static int load_balance(int this_cpu, struct rq *this_rq,
5204 struct sched_domain *sd, enum cpu_idle_type idle,
5205 int *continue_balancing)
5207 int ld_moved, cur_ld_moved, active_balance = 0;
5208 struct sched_domain *sd_parent = sd->parent;
5209 struct sched_group *group;
5211 unsigned long flags;
5212 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5214 struct lb_env env = {
5216 .dst_cpu = this_cpu,
5218 .dst_grpmask = sched_group_cpus(sd->groups),
5220 .loop_break = sched_nr_migrate_break,
5225 * For NEWLY_IDLE load_balancing, we don't need to consider
5226 * other cpus in our group
5228 if (idle == CPU_NEWLY_IDLE)
5229 env.dst_grpmask = NULL;
5231 cpumask_copy(cpus, cpu_active_mask);
5233 schedstat_inc(sd, lb_count[idle]);
5236 if (!should_we_balance(&env)) {
5237 *continue_balancing = 0;
5241 group = find_busiest_group(&env);
5243 schedstat_inc(sd, lb_nobusyg[idle]);
5247 busiest = find_busiest_queue(&env, group);
5249 schedstat_inc(sd, lb_nobusyq[idle]);
5253 BUG_ON(busiest == env.dst_rq);
5255 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5258 if (busiest->nr_running > 1) {
5260 * Attempt to move tasks. If find_busiest_group has found
5261 * an imbalance but busiest->nr_running <= 1, the group is
5262 * still unbalanced. ld_moved simply stays zero, so it is
5263 * correctly treated as an imbalance.
5265 env.flags |= LBF_ALL_PINNED;
5266 env.src_cpu = busiest->cpu;
5267 env.src_rq = busiest;
5268 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5271 local_irq_save(flags);
5272 double_rq_lock(env.dst_rq, busiest);
5275 * cur_ld_moved - load moved in current iteration
5276 * ld_moved - cumulative load moved across iterations
5278 cur_ld_moved = move_tasks(&env);
5279 ld_moved += cur_ld_moved;
5280 double_rq_unlock(env.dst_rq, busiest);
5281 local_irq_restore(flags);
5284 * some other cpu did the load balance for us.
5286 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5287 resched_cpu(env.dst_cpu);
5289 if (env.flags & LBF_NEED_BREAK) {
5290 env.flags &= ~LBF_NEED_BREAK;
5295 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5296 * us and move them to an alternate dst_cpu in our sched_group
5297 * where they can run. The upper limit on how many times we
5298 * iterate on same src_cpu is dependent on number of cpus in our
5301 * This changes load balance semantics a bit on who can move
5302 * load to a given_cpu. In addition to the given_cpu itself
5303 * (or a ilb_cpu acting on its behalf where given_cpu is
5304 * nohz-idle), we now have balance_cpu in a position to move
5305 * load to given_cpu. In rare situations, this may cause
5306 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5307 * _independently_ and at _same_ time to move some load to
5308 * given_cpu) causing exceess load to be moved to given_cpu.
5309 * This however should not happen so much in practice and
5310 * moreover subsequent load balance cycles should correct the
5311 * excess load moved.
5313 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5315 /* Prevent to re-select dst_cpu via env's cpus */
5316 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5318 env.dst_rq = cpu_rq(env.new_dst_cpu);
5319 env.dst_cpu = env.new_dst_cpu;
5320 env.flags &= ~LBF_DST_PINNED;
5322 env.loop_break = sched_nr_migrate_break;
5325 * Go back to "more_balance" rather than "redo" since we
5326 * need to continue with same src_cpu.
5332 * We failed to reach balance because of affinity.
5335 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5337 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5338 *group_imbalance = 1;
5339 } else if (*group_imbalance)
5340 *group_imbalance = 0;
5343 /* All tasks on this runqueue were pinned by CPU affinity */
5344 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5345 cpumask_clear_cpu(cpu_of(busiest), cpus);
5346 if (!cpumask_empty(cpus)) {
5348 env.loop_break = sched_nr_migrate_break;
5356 schedstat_inc(sd, lb_failed[idle]);
5358 * Increment the failure counter only on periodic balance.
5359 * We do not want newidle balance, which can be very
5360 * frequent, pollute the failure counter causing
5361 * excessive cache_hot migrations and active balances.
5363 if (idle != CPU_NEWLY_IDLE)
5364 sd->nr_balance_failed++;
5366 if (need_active_balance(&env)) {
5367 raw_spin_lock_irqsave(&busiest->lock, flags);
5369 /* don't kick the active_load_balance_cpu_stop,
5370 * if the curr task on busiest cpu can't be
5373 if (!cpumask_test_cpu(this_cpu,
5374 tsk_cpus_allowed(busiest->curr))) {
5375 raw_spin_unlock_irqrestore(&busiest->lock,
5377 env.flags |= LBF_ALL_PINNED;
5378 goto out_one_pinned;
5382 * ->active_balance synchronizes accesses to
5383 * ->active_balance_work. Once set, it's cleared
5384 * only after active load balance is finished.
5386 if (!busiest->active_balance) {
5387 busiest->active_balance = 1;
5388 busiest->push_cpu = this_cpu;
5391 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5393 if (active_balance) {
5394 stop_one_cpu_nowait(cpu_of(busiest),
5395 active_load_balance_cpu_stop, busiest,
5396 &busiest->active_balance_work);
5400 * We've kicked active balancing, reset the failure
5403 sd->nr_balance_failed = sd->cache_nice_tries+1;
5406 sd->nr_balance_failed = 0;
5408 if (likely(!active_balance)) {
5409 /* We were unbalanced, so reset the balancing interval */
5410 sd->balance_interval = sd->min_interval;
5413 * If we've begun active balancing, start to back off. This
5414 * case may not be covered by the all_pinned logic if there
5415 * is only 1 task on the busy runqueue (because we don't call
5418 if (sd->balance_interval < sd->max_interval)
5419 sd->balance_interval *= 2;
5425 schedstat_inc(sd, lb_balanced[idle]);
5427 sd->nr_balance_failed = 0;
5430 /* tune up the balancing interval */
5431 if (((env.flags & LBF_ALL_PINNED) &&
5432 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5433 (sd->balance_interval < sd->max_interval))
5434 sd->balance_interval *= 2;
5442 * idle_balance is called by schedule() if this_cpu is about to become
5443 * idle. Attempts to pull tasks from other CPUs.
5445 void idle_balance(int this_cpu, struct rq *this_rq)
5447 struct sched_domain *sd;
5448 int pulled_task = 0;
5449 unsigned long next_balance = jiffies + HZ;
5452 this_rq->idle_stamp = rq_clock(this_rq);
5454 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5458 * Drop the rq->lock, but keep IRQ/preempt disabled.
5460 raw_spin_unlock(&this_rq->lock);
5462 update_blocked_averages(this_cpu);
5464 for_each_domain(this_cpu, sd) {
5465 unsigned long interval;
5466 int continue_balancing = 1;
5467 u64 t0, domain_cost;
5469 if (!(sd->flags & SD_LOAD_BALANCE))
5472 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5475 if (sd->flags & SD_BALANCE_NEWIDLE) {
5476 t0 = sched_clock_cpu(this_cpu);
5478 /* If we've pulled tasks over stop searching: */
5479 pulled_task = load_balance(this_cpu, this_rq,
5481 &continue_balancing);
5483 domain_cost = sched_clock_cpu(this_cpu) - t0;
5484 if (domain_cost > sd->max_newidle_lb_cost)
5485 sd->max_newidle_lb_cost = domain_cost;
5487 curr_cost += domain_cost;
5490 interval = msecs_to_jiffies(sd->balance_interval);
5491 if (time_after(next_balance, sd->last_balance + interval))
5492 next_balance = sd->last_balance + interval;
5494 this_rq->idle_stamp = 0;
5500 raw_spin_lock(&this_rq->lock);
5502 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5504 * We are going idle. next_balance may be set based on
5505 * a busy processor. So reset next_balance.
5507 this_rq->next_balance = next_balance;
5510 if (curr_cost > this_rq->max_idle_balance_cost)
5511 this_rq->max_idle_balance_cost = curr_cost;
5515 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5516 * running tasks off the busiest CPU onto idle CPUs. It requires at
5517 * least 1 task to be running on each physical CPU where possible, and
5518 * avoids physical / logical imbalances.
5520 static int active_load_balance_cpu_stop(void *data)
5522 struct rq *busiest_rq = data;
5523 int busiest_cpu = cpu_of(busiest_rq);
5524 int target_cpu = busiest_rq->push_cpu;
5525 struct rq *target_rq = cpu_rq(target_cpu);
5526 struct sched_domain *sd;
5528 raw_spin_lock_irq(&busiest_rq->lock);
5530 /* make sure the requested cpu hasn't gone down in the meantime */
5531 if (unlikely(busiest_cpu != smp_processor_id() ||
5532 !busiest_rq->active_balance))
5535 /* Is there any task to move? */
5536 if (busiest_rq->nr_running <= 1)
5540 * This condition is "impossible", if it occurs
5541 * we need to fix it. Originally reported by
5542 * Bjorn Helgaas on a 128-cpu setup.
5544 BUG_ON(busiest_rq == target_rq);
5546 /* move a task from busiest_rq to target_rq */
5547 double_lock_balance(busiest_rq, target_rq);
5549 /* Search for an sd spanning us and the target CPU. */
5551 for_each_domain(target_cpu, sd) {
5552 if ((sd->flags & SD_LOAD_BALANCE) &&
5553 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5558 struct lb_env env = {
5560 .dst_cpu = target_cpu,
5561 .dst_rq = target_rq,
5562 .src_cpu = busiest_rq->cpu,
5563 .src_rq = busiest_rq,
5567 schedstat_inc(sd, alb_count);
5569 if (move_one_task(&env))
5570 schedstat_inc(sd, alb_pushed);
5572 schedstat_inc(sd, alb_failed);
5575 double_unlock_balance(busiest_rq, target_rq);
5577 busiest_rq->active_balance = 0;
5578 raw_spin_unlock_irq(&busiest_rq->lock);
5582 #ifdef CONFIG_NO_HZ_COMMON
5584 * idle load balancing details
5585 * - When one of the busy CPUs notice that there may be an idle rebalancing
5586 * needed, they will kick the idle load balancer, which then does idle
5587 * load balancing for all the idle CPUs.
5590 cpumask_var_t idle_cpus_mask;
5592 unsigned long next_balance; /* in jiffy units */
5593 } nohz ____cacheline_aligned;
5595 static inline int find_new_ilb(int call_cpu)
5597 int ilb = cpumask_first(nohz.idle_cpus_mask);
5599 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5606 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5607 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5608 * CPU (if there is one).
5610 static void nohz_balancer_kick(int cpu)
5614 nohz.next_balance++;
5616 ilb_cpu = find_new_ilb(cpu);
5618 if (ilb_cpu >= nr_cpu_ids)
5621 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5624 * Use smp_send_reschedule() instead of resched_cpu().
5625 * This way we generate a sched IPI on the target cpu which
5626 * is idle. And the softirq performing nohz idle load balance
5627 * will be run before returning from the IPI.
5629 smp_send_reschedule(ilb_cpu);
5633 static inline void nohz_balance_exit_idle(int cpu)
5635 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5636 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5637 atomic_dec(&nohz.nr_cpus);
5638 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5642 static inline void set_cpu_sd_state_busy(void)
5644 struct sched_domain *sd;
5647 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5649 if (!sd || !sd->nohz_idle)
5653 for (; sd; sd = sd->parent)
5654 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5659 void set_cpu_sd_state_idle(void)
5661 struct sched_domain *sd;
5664 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5666 if (!sd || sd->nohz_idle)
5670 for (; sd; sd = sd->parent)
5671 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5677 * This routine will record that the cpu is going idle with tick stopped.
5678 * This info will be used in performing idle load balancing in the future.
5680 void nohz_balance_enter_idle(int cpu)
5683 * If this cpu is going down, then nothing needs to be done.
5685 if (!cpu_active(cpu))
5688 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5691 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5692 atomic_inc(&nohz.nr_cpus);
5693 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5696 static int sched_ilb_notifier(struct notifier_block *nfb,
5697 unsigned long action, void *hcpu)
5699 switch (action & ~CPU_TASKS_FROZEN) {
5701 nohz_balance_exit_idle(smp_processor_id());
5709 static DEFINE_SPINLOCK(balancing);
5712 * Scale the max load_balance interval with the number of CPUs in the system.
5713 * This trades load-balance latency on larger machines for less cross talk.
5715 void update_max_interval(void)
5717 max_load_balance_interval = HZ*num_online_cpus()/10;
5721 * It checks each scheduling domain to see if it is due to be balanced,
5722 * and initiates a balancing operation if so.
5724 * Balancing parameters are set up in init_sched_domains.
5726 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5728 int continue_balancing = 1;
5729 struct rq *rq = cpu_rq(cpu);
5730 unsigned long interval;
5731 struct sched_domain *sd;
5732 /* Earliest time when we have to do rebalance again */
5733 unsigned long next_balance = jiffies + 60*HZ;
5734 int update_next_balance = 0;
5735 int need_serialize, need_decay = 0;
5738 update_blocked_averages(cpu);
5741 for_each_domain(cpu, sd) {
5743 * Decay the newidle max times here because this is a regular
5744 * visit to all the domains. Decay ~1% per second.
5746 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5747 sd->max_newidle_lb_cost =
5748 (sd->max_newidle_lb_cost * 253) / 256;
5749 sd->next_decay_max_lb_cost = jiffies + HZ;
5752 max_cost += sd->max_newidle_lb_cost;
5754 if (!(sd->flags & SD_LOAD_BALANCE))
5758 * Stop the load balance at this level. There is another
5759 * CPU in our sched group which is doing load balancing more
5762 if (!continue_balancing) {
5768 interval = sd->balance_interval;
5769 if (idle != CPU_IDLE)
5770 interval *= sd->busy_factor;
5772 /* scale ms to jiffies */
5773 interval = msecs_to_jiffies(interval);
5774 interval = clamp(interval, 1UL, max_load_balance_interval);
5776 need_serialize = sd->flags & SD_SERIALIZE;
5778 if (need_serialize) {
5779 if (!spin_trylock(&balancing))
5783 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5784 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5786 * The LBF_DST_PINNED logic could have changed
5787 * env->dst_cpu, so we can't know our idle
5788 * state even if we migrated tasks. Update it.
5790 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5792 sd->last_balance = jiffies;
5795 spin_unlock(&balancing);
5797 if (time_after(next_balance, sd->last_balance + interval)) {
5798 next_balance = sd->last_balance + interval;
5799 update_next_balance = 1;
5804 * Ensure the rq-wide value also decays but keep it at a
5805 * reasonable floor to avoid funnies with rq->avg_idle.
5807 rq->max_idle_balance_cost =
5808 max((u64)sysctl_sched_migration_cost, max_cost);
5813 * next_balance will be updated only when there is a need.
5814 * When the cpu is attached to null domain for ex, it will not be
5817 if (likely(update_next_balance))
5818 rq->next_balance = next_balance;
5821 #ifdef CONFIG_NO_HZ_COMMON
5823 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5824 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5826 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5828 struct rq *this_rq = cpu_rq(this_cpu);
5832 if (idle != CPU_IDLE ||
5833 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5836 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5837 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5841 * If this cpu gets work to do, stop the load balancing
5842 * work being done for other cpus. Next load
5843 * balancing owner will pick it up.
5848 rq = cpu_rq(balance_cpu);
5850 raw_spin_lock_irq(&rq->lock);
5851 update_rq_clock(rq);
5852 update_idle_cpu_load(rq);
5853 raw_spin_unlock_irq(&rq->lock);
5855 rebalance_domains(balance_cpu, CPU_IDLE);
5857 if (time_after(this_rq->next_balance, rq->next_balance))
5858 this_rq->next_balance = rq->next_balance;
5860 nohz.next_balance = this_rq->next_balance;
5862 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5866 * Current heuristic for kicking the idle load balancer in the presence
5867 * of an idle cpu is the system.
5868 * - This rq has more than one task.
5869 * - At any scheduler domain level, this cpu's scheduler group has multiple
5870 * busy cpu's exceeding the group's power.
5871 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5872 * domain span are idle.
5874 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5876 unsigned long now = jiffies;
5877 struct sched_domain *sd;
5879 if (unlikely(idle_cpu(cpu)))
5883 * We may be recently in ticked or tickless idle mode. At the first
5884 * busy tick after returning from idle, we will update the busy stats.
5886 set_cpu_sd_state_busy();
5887 nohz_balance_exit_idle(cpu);
5890 * None are in tickless mode and hence no need for NOHZ idle load
5893 if (likely(!atomic_read(&nohz.nr_cpus)))
5896 if (time_before(now, nohz.next_balance))
5899 if (rq->nr_running >= 2)
5903 for_each_domain(cpu, sd) {
5904 struct sched_group *sg = sd->groups;
5905 struct sched_group_power *sgp = sg->sgp;
5906 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5908 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5909 goto need_kick_unlock;
5911 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5912 && (cpumask_first_and(nohz.idle_cpus_mask,
5913 sched_domain_span(sd)) < cpu))
5914 goto need_kick_unlock;
5916 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5928 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5932 * run_rebalance_domains is triggered when needed from the scheduler tick.
5933 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5935 static void run_rebalance_domains(struct softirq_action *h)
5937 int this_cpu = smp_processor_id();
5938 struct rq *this_rq = cpu_rq(this_cpu);
5939 enum cpu_idle_type idle = this_rq->idle_balance ?
5940 CPU_IDLE : CPU_NOT_IDLE;
5942 rebalance_domains(this_cpu, idle);
5945 * If this cpu has a pending nohz_balance_kick, then do the
5946 * balancing on behalf of the other idle cpus whose ticks are
5949 nohz_idle_balance(this_cpu, idle);
5952 static inline int on_null_domain(int cpu)
5954 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5958 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5960 void trigger_load_balance(struct rq *rq, int cpu)
5962 /* Don't need to rebalance while attached to NULL domain */
5963 if (time_after_eq(jiffies, rq->next_balance) &&
5964 likely(!on_null_domain(cpu)))
5965 raise_softirq(SCHED_SOFTIRQ);
5966 #ifdef CONFIG_NO_HZ_COMMON
5967 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5968 nohz_balancer_kick(cpu);
5972 static void rq_online_fair(struct rq *rq)
5977 static void rq_offline_fair(struct rq *rq)
5981 /* Ensure any throttled groups are reachable by pick_next_task */
5982 unthrottle_offline_cfs_rqs(rq);
5985 #endif /* CONFIG_SMP */
5988 * scheduler tick hitting a task of our scheduling class:
5990 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5992 struct cfs_rq *cfs_rq;
5993 struct sched_entity *se = &curr->se;
5995 for_each_sched_entity(se) {
5996 cfs_rq = cfs_rq_of(se);
5997 entity_tick(cfs_rq, se, queued);
6000 if (numabalancing_enabled)
6001 task_tick_numa(rq, curr);
6003 update_rq_runnable_avg(rq, 1);
6007 * called on fork with the child task as argument from the parent's context
6008 * - child not yet on the tasklist
6009 * - preemption disabled
6011 static void task_fork_fair(struct task_struct *p)
6013 struct cfs_rq *cfs_rq;
6014 struct sched_entity *se = &p->se, *curr;
6015 int this_cpu = smp_processor_id();
6016 struct rq *rq = this_rq();
6017 unsigned long flags;
6019 raw_spin_lock_irqsave(&rq->lock, flags);
6021 update_rq_clock(rq);
6023 cfs_rq = task_cfs_rq(current);
6024 curr = cfs_rq->curr;
6027 * Not only the cpu but also the task_group of the parent might have
6028 * been changed after parent->se.parent,cfs_rq were copied to
6029 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6030 * of child point to valid ones.
6033 __set_task_cpu(p, this_cpu);
6036 update_curr(cfs_rq);
6039 se->vruntime = curr->vruntime;
6040 place_entity(cfs_rq, se, 1);
6042 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6044 * Upon rescheduling, sched_class::put_prev_task() will place
6045 * 'current' within the tree based on its new key value.
6047 swap(curr->vruntime, se->vruntime);
6048 resched_task(rq->curr);
6051 se->vruntime -= cfs_rq->min_vruntime;
6053 raw_spin_unlock_irqrestore(&rq->lock, flags);
6057 * Priority of the task has changed. Check to see if we preempt
6061 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6067 * Reschedule if we are currently running on this runqueue and
6068 * our priority decreased, or if we are not currently running on
6069 * this runqueue and our priority is higher than the current's
6071 if (rq->curr == p) {
6072 if (p->prio > oldprio)
6073 resched_task(rq->curr);
6075 check_preempt_curr(rq, p, 0);
6078 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6080 struct sched_entity *se = &p->se;
6081 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6084 * Ensure the task's vruntime is normalized, so that when its
6085 * switched back to the fair class the enqueue_entity(.flags=0) will
6086 * do the right thing.
6088 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6089 * have normalized the vruntime, if it was !on_rq, then only when
6090 * the task is sleeping will it still have non-normalized vruntime.
6092 if (!se->on_rq && p->state != TASK_RUNNING) {
6094 * Fix up our vruntime so that the current sleep doesn't
6095 * cause 'unlimited' sleep bonus.
6097 place_entity(cfs_rq, se, 0);
6098 se->vruntime -= cfs_rq->min_vruntime;
6103 * Remove our load from contribution when we leave sched_fair
6104 * and ensure we don't carry in an old decay_count if we
6107 if (se->avg.decay_count) {
6108 __synchronize_entity_decay(se);
6109 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6115 * We switched to the sched_fair class.
6117 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6123 * We were most likely switched from sched_rt, so
6124 * kick off the schedule if running, otherwise just see
6125 * if we can still preempt the current task.
6128 resched_task(rq->curr);
6130 check_preempt_curr(rq, p, 0);
6133 /* Account for a task changing its policy or group.
6135 * This routine is mostly called to set cfs_rq->curr field when a task
6136 * migrates between groups/classes.
6138 static void set_curr_task_fair(struct rq *rq)
6140 struct sched_entity *se = &rq->curr->se;
6142 for_each_sched_entity(se) {
6143 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6145 set_next_entity(cfs_rq, se);
6146 /* ensure bandwidth has been allocated on our new cfs_rq */
6147 account_cfs_rq_runtime(cfs_rq, 0);
6151 void init_cfs_rq(struct cfs_rq *cfs_rq)
6153 cfs_rq->tasks_timeline = RB_ROOT;
6154 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6155 #ifndef CONFIG_64BIT
6156 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6159 atomic64_set(&cfs_rq->decay_counter, 1);
6160 atomic_long_set(&cfs_rq->removed_load, 0);
6164 #ifdef CONFIG_FAIR_GROUP_SCHED
6165 static void task_move_group_fair(struct task_struct *p, int on_rq)
6167 struct cfs_rq *cfs_rq;
6169 * If the task was not on the rq at the time of this cgroup movement
6170 * it must have been asleep, sleeping tasks keep their ->vruntime
6171 * absolute on their old rq until wakeup (needed for the fair sleeper
6172 * bonus in place_entity()).
6174 * If it was on the rq, we've just 'preempted' it, which does convert
6175 * ->vruntime to a relative base.
6177 * Make sure both cases convert their relative position when migrating
6178 * to another cgroup's rq. This does somewhat interfere with the
6179 * fair sleeper stuff for the first placement, but who cares.
6182 * When !on_rq, vruntime of the task has usually NOT been normalized.
6183 * But there are some cases where it has already been normalized:
6185 * - Moving a forked child which is waiting for being woken up by
6186 * wake_up_new_task().
6187 * - Moving a task which has been woken up by try_to_wake_up() and
6188 * waiting for actually being woken up by sched_ttwu_pending().
6190 * To prevent boost or penalty in the new cfs_rq caused by delta
6191 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6193 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6197 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6198 set_task_rq(p, task_cpu(p));
6200 cfs_rq = cfs_rq_of(&p->se);
6201 p->se.vruntime += cfs_rq->min_vruntime;
6204 * migrate_task_rq_fair() will have removed our previous
6205 * contribution, but we must synchronize for ongoing future
6208 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6209 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6214 void free_fair_sched_group(struct task_group *tg)
6218 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6220 for_each_possible_cpu(i) {
6222 kfree(tg->cfs_rq[i]);
6231 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6233 struct cfs_rq *cfs_rq;
6234 struct sched_entity *se;
6237 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6240 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6244 tg->shares = NICE_0_LOAD;
6246 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6248 for_each_possible_cpu(i) {
6249 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6250 GFP_KERNEL, cpu_to_node(i));
6254 se = kzalloc_node(sizeof(struct sched_entity),
6255 GFP_KERNEL, cpu_to_node(i));
6259 init_cfs_rq(cfs_rq);
6260 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6271 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6273 struct rq *rq = cpu_rq(cpu);
6274 unsigned long flags;
6277 * Only empty task groups can be destroyed; so we can speculatively
6278 * check on_list without danger of it being re-added.
6280 if (!tg->cfs_rq[cpu]->on_list)
6283 raw_spin_lock_irqsave(&rq->lock, flags);
6284 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6285 raw_spin_unlock_irqrestore(&rq->lock, flags);
6288 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6289 struct sched_entity *se, int cpu,
6290 struct sched_entity *parent)
6292 struct rq *rq = cpu_rq(cpu);
6296 init_cfs_rq_runtime(cfs_rq);
6298 tg->cfs_rq[cpu] = cfs_rq;
6301 /* se could be NULL for root_task_group */
6306 se->cfs_rq = &rq->cfs;
6308 se->cfs_rq = parent->my_q;
6311 update_load_set(&se->load, 0);
6312 se->parent = parent;
6315 static DEFINE_MUTEX(shares_mutex);
6317 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6320 unsigned long flags;
6323 * We can't change the weight of the root cgroup.
6328 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6330 mutex_lock(&shares_mutex);
6331 if (tg->shares == shares)
6334 tg->shares = shares;
6335 for_each_possible_cpu(i) {
6336 struct rq *rq = cpu_rq(i);
6337 struct sched_entity *se;
6340 /* Propagate contribution to hierarchy */
6341 raw_spin_lock_irqsave(&rq->lock, flags);
6343 /* Possible calls to update_curr() need rq clock */
6344 update_rq_clock(rq);
6345 for_each_sched_entity(se)
6346 update_cfs_shares(group_cfs_rq(se));
6347 raw_spin_unlock_irqrestore(&rq->lock, flags);
6351 mutex_unlock(&shares_mutex);
6354 #else /* CONFIG_FAIR_GROUP_SCHED */
6356 void free_fair_sched_group(struct task_group *tg) { }
6358 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6363 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6365 #endif /* CONFIG_FAIR_GROUP_SCHED */
6368 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6370 struct sched_entity *se = &task->se;
6371 unsigned int rr_interval = 0;
6374 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6377 if (rq->cfs.load.weight)
6378 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6384 * All the scheduling class methods:
6386 const struct sched_class fair_sched_class = {
6387 .next = &idle_sched_class,
6388 .enqueue_task = enqueue_task_fair,
6389 .dequeue_task = dequeue_task_fair,
6390 .yield_task = yield_task_fair,
6391 .yield_to_task = yield_to_task_fair,
6393 .check_preempt_curr = check_preempt_wakeup,
6395 .pick_next_task = pick_next_task_fair,
6396 .put_prev_task = put_prev_task_fair,
6399 .select_task_rq = select_task_rq_fair,
6400 .migrate_task_rq = migrate_task_rq_fair,
6402 .rq_online = rq_online_fair,
6403 .rq_offline = rq_offline_fair,
6405 .task_waking = task_waking_fair,
6408 .set_curr_task = set_curr_task_fair,
6409 .task_tick = task_tick_fair,
6410 .task_fork = task_fork_fair,
6412 .prio_changed = prio_changed_fair,
6413 .switched_from = switched_from_fair,
6414 .switched_to = switched_to_fair,
6416 .get_rr_interval = get_rr_interval_fair,
6418 #ifdef CONFIG_FAIR_GROUP_SCHED
6419 .task_move_group = task_move_group_fair,
6423 #ifdef CONFIG_SCHED_DEBUG
6424 void print_cfs_stats(struct seq_file *m, int cpu)
6426 struct cfs_rq *cfs_rq;
6429 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6430 print_cfs_rq(m, cpu, cfs_rq);
6435 __init void init_sched_fair_class(void)
6438 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6440 #ifdef CONFIG_NO_HZ_COMMON
6441 nohz.next_balance = jiffies;
6442 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6443 cpu_notifier(sched_ilb_notifier, 0);