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)
882 int seq, nid, max_nid = -1;
883 unsigned long max_faults = 0;
885 if (!p->mm) /* for example, ksmd faulting in a user's mm */
887 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
888 if (p->numa_scan_seq == seq)
890 p->numa_scan_seq = seq;
891 p->numa_scan_period_max = task_scan_max(p);
893 /* Find the node with the highest number of faults */
894 for_each_online_node(nid) {
895 unsigned long faults = p->numa_faults[nid];
896 p->numa_faults[nid] >>= 1;
897 if (faults > max_faults) {
903 /* Update the tasks preferred node if necessary */
904 if (max_faults && max_nid != p->numa_preferred_nid)
905 p->numa_preferred_nid = max_nid;
909 * Got a PROT_NONE fault for a page on @node.
911 void task_numa_fault(int node, int pages, bool migrated)
913 struct task_struct *p = current;
915 if (!numabalancing_enabled)
918 /* Allocate buffer to track faults on a per-node basis */
919 if (unlikely(!p->numa_faults)) {
920 int size = sizeof(*p->numa_faults) * nr_node_ids;
922 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
928 * If pages are properly placed (did not migrate) then scan slower.
929 * This is reset periodically in case of phase changes
932 /* Initialise if necessary */
933 if (!p->numa_scan_period_max)
934 p->numa_scan_period_max = task_scan_max(p);
936 p->numa_scan_period = min(p->numa_scan_period_max,
937 p->numa_scan_period + 10);
940 task_numa_placement(p);
942 p->numa_faults[node] += pages;
945 static void reset_ptenuma_scan(struct task_struct *p)
947 ACCESS_ONCE(p->mm->numa_scan_seq)++;
948 p->mm->numa_scan_offset = 0;
952 * The expensive part of numa migration is done from task_work context.
953 * Triggered from task_tick_numa().
955 void task_numa_work(struct callback_head *work)
957 unsigned long migrate, next_scan, now = jiffies;
958 struct task_struct *p = current;
959 struct mm_struct *mm = p->mm;
960 struct vm_area_struct *vma;
961 unsigned long start, end;
962 unsigned long nr_pte_updates = 0;
965 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
967 work->next = work; /* protect against double add */
969 * Who cares about NUMA placement when they're dying.
971 * NOTE: make sure not to dereference p->mm before this check,
972 * exit_task_work() happens _after_ exit_mm() so we could be called
973 * without p->mm even though we still had it when we enqueued this
976 if (p->flags & PF_EXITING)
979 if (!mm->numa_next_reset || !mm->numa_next_scan) {
980 mm->numa_next_scan = now +
981 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
982 mm->numa_next_reset = now +
983 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
987 * Reset the scan period if enough time has gone by. Objective is that
988 * scanning will be reduced if pages are properly placed. As tasks
989 * can enter different phases this needs to be re-examined. Lacking
990 * proper tracking of reference behaviour, this blunt hammer is used.
992 migrate = mm->numa_next_reset;
993 if (time_after(now, migrate)) {
994 p->numa_scan_period = task_scan_min(p);
995 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
996 xchg(&mm->numa_next_reset, next_scan);
1000 * Enforce maximal scan/migration frequency..
1002 migrate = mm->numa_next_scan;
1003 if (time_before(now, migrate))
1006 if (p->numa_scan_period == 0) {
1007 p->numa_scan_period_max = task_scan_max(p);
1008 p->numa_scan_period = task_scan_min(p);
1011 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1012 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1016 * Delay this task enough that another task of this mm will likely win
1017 * the next time around.
1019 p->node_stamp += 2 * TICK_NSEC;
1021 start = mm->numa_scan_offset;
1022 pages = sysctl_numa_balancing_scan_size;
1023 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1027 down_read(&mm->mmap_sem);
1028 vma = find_vma(mm, start);
1030 reset_ptenuma_scan(p);
1034 for (; vma; vma = vma->vm_next) {
1035 if (!vma_migratable(vma))
1038 /* Skip small VMAs. They are not likely to be of relevance */
1039 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1043 start = max(start, vma->vm_start);
1044 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1045 end = min(end, vma->vm_end);
1046 nr_pte_updates += change_prot_numa(vma, start, end);
1049 * Scan sysctl_numa_balancing_scan_size but ensure that
1050 * at least one PTE is updated so that unused virtual
1051 * address space is quickly skipped.
1054 pages -= (end - start) >> PAGE_SHIFT;
1059 } while (end != vma->vm_end);
1064 * If the whole process was scanned without updates then no NUMA
1065 * hinting faults are being recorded and scan rate should be lower.
1067 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1068 p->numa_scan_period = min(p->numa_scan_period_max,
1069 p->numa_scan_period << 1);
1071 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1072 mm->numa_next_scan = next_scan;
1076 * It is possible to reach the end of the VMA list but the last few
1077 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1078 * would find the !migratable VMA on the next scan but not reset the
1079 * scanner to the start so check it now.
1082 mm->numa_scan_offset = start;
1084 reset_ptenuma_scan(p);
1085 up_read(&mm->mmap_sem);
1089 * Drive the periodic memory faults..
1091 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1093 struct callback_head *work = &curr->numa_work;
1097 * We don't care about NUMA placement if we don't have memory.
1099 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1103 * Using runtime rather than walltime has the dual advantage that
1104 * we (mostly) drive the selection from busy threads and that the
1105 * task needs to have done some actual work before we bother with
1108 now = curr->se.sum_exec_runtime;
1109 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1111 if (now - curr->node_stamp > period) {
1112 if (!curr->node_stamp)
1113 curr->numa_scan_period = task_scan_min(curr);
1114 curr->node_stamp += period;
1116 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1117 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1118 task_work_add(curr, work, true);
1123 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1126 #endif /* CONFIG_NUMA_BALANCING */
1129 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1131 update_load_add(&cfs_rq->load, se->load.weight);
1132 if (!parent_entity(se))
1133 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1135 if (entity_is_task(se))
1136 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1138 cfs_rq->nr_running++;
1142 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1144 update_load_sub(&cfs_rq->load, se->load.weight);
1145 if (!parent_entity(se))
1146 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1147 if (entity_is_task(se))
1148 list_del_init(&se->group_node);
1149 cfs_rq->nr_running--;
1152 #ifdef CONFIG_FAIR_GROUP_SCHED
1154 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1159 * Use this CPU's actual weight instead of the last load_contribution
1160 * to gain a more accurate current total weight. See
1161 * update_cfs_rq_load_contribution().
1163 tg_weight = atomic_long_read(&tg->load_avg);
1164 tg_weight -= cfs_rq->tg_load_contrib;
1165 tg_weight += cfs_rq->load.weight;
1170 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1172 long tg_weight, load, shares;
1174 tg_weight = calc_tg_weight(tg, cfs_rq);
1175 load = cfs_rq->load.weight;
1177 shares = (tg->shares * load);
1179 shares /= tg_weight;
1181 if (shares < MIN_SHARES)
1182 shares = MIN_SHARES;
1183 if (shares > tg->shares)
1184 shares = tg->shares;
1188 # else /* CONFIG_SMP */
1189 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1193 # endif /* CONFIG_SMP */
1194 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1195 unsigned long weight)
1198 /* commit outstanding execution time */
1199 if (cfs_rq->curr == se)
1200 update_curr(cfs_rq);
1201 account_entity_dequeue(cfs_rq, se);
1204 update_load_set(&se->load, weight);
1207 account_entity_enqueue(cfs_rq, se);
1210 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1212 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1214 struct task_group *tg;
1215 struct sched_entity *se;
1219 se = tg->se[cpu_of(rq_of(cfs_rq))];
1220 if (!se || throttled_hierarchy(cfs_rq))
1223 if (likely(se->load.weight == tg->shares))
1226 shares = calc_cfs_shares(cfs_rq, tg);
1228 reweight_entity(cfs_rq_of(se), se, shares);
1230 #else /* CONFIG_FAIR_GROUP_SCHED */
1231 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1234 #endif /* CONFIG_FAIR_GROUP_SCHED */
1238 * We choose a half-life close to 1 scheduling period.
1239 * Note: The tables below are dependent on this value.
1241 #define LOAD_AVG_PERIOD 32
1242 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1243 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1245 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1246 static const u32 runnable_avg_yN_inv[] = {
1247 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1248 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1249 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1250 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1251 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1252 0x85aac367, 0x82cd8698,
1256 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1257 * over-estimates when re-combining.
1259 static const u32 runnable_avg_yN_sum[] = {
1260 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1261 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1262 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1267 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1269 static __always_inline u64 decay_load(u64 val, u64 n)
1271 unsigned int local_n;
1275 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1278 /* after bounds checking we can collapse to 32-bit */
1282 * As y^PERIOD = 1/2, we can combine
1283 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1284 * With a look-up table which covers k^n (n<PERIOD)
1286 * To achieve constant time decay_load.
1288 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1289 val >>= local_n / LOAD_AVG_PERIOD;
1290 local_n %= LOAD_AVG_PERIOD;
1293 val *= runnable_avg_yN_inv[local_n];
1294 /* We don't use SRR here since we always want to round down. */
1299 * For updates fully spanning n periods, the contribution to runnable
1300 * average will be: \Sum 1024*y^n
1302 * We can compute this reasonably efficiently by combining:
1303 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1305 static u32 __compute_runnable_contrib(u64 n)
1309 if (likely(n <= LOAD_AVG_PERIOD))
1310 return runnable_avg_yN_sum[n];
1311 else if (unlikely(n >= LOAD_AVG_MAX_N))
1312 return LOAD_AVG_MAX;
1314 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1316 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1317 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1319 n -= LOAD_AVG_PERIOD;
1320 } while (n > LOAD_AVG_PERIOD);
1322 contrib = decay_load(contrib, n);
1323 return contrib + runnable_avg_yN_sum[n];
1327 * We can represent the historical contribution to runnable average as the
1328 * coefficients of a geometric series. To do this we sub-divide our runnable
1329 * history into segments of approximately 1ms (1024us); label the segment that
1330 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1332 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1334 * (now) (~1ms ago) (~2ms ago)
1336 * Let u_i denote the fraction of p_i that the entity was runnable.
1338 * We then designate the fractions u_i as our co-efficients, yielding the
1339 * following representation of historical load:
1340 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1342 * We choose y based on the with of a reasonably scheduling period, fixing:
1345 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1346 * approximately half as much as the contribution to load within the last ms
1349 * When a period "rolls over" and we have new u_0`, multiplying the previous
1350 * sum again by y is sufficient to update:
1351 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1352 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1354 static __always_inline int __update_entity_runnable_avg(u64 now,
1355 struct sched_avg *sa,
1359 u32 runnable_contrib;
1360 int delta_w, decayed = 0;
1362 delta = now - sa->last_runnable_update;
1364 * This should only happen when time goes backwards, which it
1365 * unfortunately does during sched clock init when we swap over to TSC.
1367 if ((s64)delta < 0) {
1368 sa->last_runnable_update = now;
1373 * Use 1024ns as the unit of measurement since it's a reasonable
1374 * approximation of 1us and fast to compute.
1379 sa->last_runnable_update = now;
1381 /* delta_w is the amount already accumulated against our next period */
1382 delta_w = sa->runnable_avg_period % 1024;
1383 if (delta + delta_w >= 1024) {
1384 /* period roll-over */
1388 * Now that we know we're crossing a period boundary, figure
1389 * out how much from delta we need to complete the current
1390 * period and accrue it.
1392 delta_w = 1024 - delta_w;
1394 sa->runnable_avg_sum += delta_w;
1395 sa->runnable_avg_period += delta_w;
1399 /* Figure out how many additional periods this update spans */
1400 periods = delta / 1024;
1403 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1405 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1408 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1409 runnable_contrib = __compute_runnable_contrib(periods);
1411 sa->runnable_avg_sum += runnable_contrib;
1412 sa->runnable_avg_period += runnable_contrib;
1415 /* Remainder of delta accrued against u_0` */
1417 sa->runnable_avg_sum += delta;
1418 sa->runnable_avg_period += delta;
1423 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1424 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1426 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1427 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1429 decays -= se->avg.decay_count;
1433 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1434 se->avg.decay_count = 0;
1439 #ifdef CONFIG_FAIR_GROUP_SCHED
1440 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1443 struct task_group *tg = cfs_rq->tg;
1446 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1447 tg_contrib -= cfs_rq->tg_load_contrib;
1449 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1450 atomic_long_add(tg_contrib, &tg->load_avg);
1451 cfs_rq->tg_load_contrib += tg_contrib;
1456 * Aggregate cfs_rq runnable averages into an equivalent task_group
1457 * representation for computing load contributions.
1459 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1460 struct cfs_rq *cfs_rq)
1462 struct task_group *tg = cfs_rq->tg;
1465 /* The fraction of a cpu used by this cfs_rq */
1466 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1467 sa->runnable_avg_period + 1);
1468 contrib -= cfs_rq->tg_runnable_contrib;
1470 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1471 atomic_add(contrib, &tg->runnable_avg);
1472 cfs_rq->tg_runnable_contrib += contrib;
1476 static inline void __update_group_entity_contrib(struct sched_entity *se)
1478 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1479 struct task_group *tg = cfs_rq->tg;
1484 contrib = cfs_rq->tg_load_contrib * tg->shares;
1485 se->avg.load_avg_contrib = div_u64(contrib,
1486 atomic_long_read(&tg->load_avg) + 1);
1489 * For group entities we need to compute a correction term in the case
1490 * that they are consuming <1 cpu so that we would contribute the same
1491 * load as a task of equal weight.
1493 * Explicitly co-ordinating this measurement would be expensive, but
1494 * fortunately the sum of each cpus contribution forms a usable
1495 * lower-bound on the true value.
1497 * Consider the aggregate of 2 contributions. Either they are disjoint
1498 * (and the sum represents true value) or they are disjoint and we are
1499 * understating by the aggregate of their overlap.
1501 * Extending this to N cpus, for a given overlap, the maximum amount we
1502 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1503 * cpus that overlap for this interval and w_i is the interval width.
1505 * On a small machine; the first term is well-bounded which bounds the
1506 * total error since w_i is a subset of the period. Whereas on a
1507 * larger machine, while this first term can be larger, if w_i is the
1508 * of consequential size guaranteed to see n_i*w_i quickly converge to
1509 * our upper bound of 1-cpu.
1511 runnable_avg = atomic_read(&tg->runnable_avg);
1512 if (runnable_avg < NICE_0_LOAD) {
1513 se->avg.load_avg_contrib *= runnable_avg;
1514 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1518 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1519 int force_update) {}
1520 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1521 struct cfs_rq *cfs_rq) {}
1522 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1525 static inline void __update_task_entity_contrib(struct sched_entity *se)
1529 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1530 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1531 contrib /= (se->avg.runnable_avg_period + 1);
1532 se->avg.load_avg_contrib = scale_load(contrib);
1535 /* Compute the current contribution to load_avg by se, return any delta */
1536 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1538 long old_contrib = se->avg.load_avg_contrib;
1540 if (entity_is_task(se)) {
1541 __update_task_entity_contrib(se);
1543 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1544 __update_group_entity_contrib(se);
1547 return se->avg.load_avg_contrib - old_contrib;
1550 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1553 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1554 cfs_rq->blocked_load_avg -= load_contrib;
1556 cfs_rq->blocked_load_avg = 0;
1559 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1561 /* Update a sched_entity's runnable average */
1562 static inline void update_entity_load_avg(struct sched_entity *se,
1565 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1570 * For a group entity we need to use their owned cfs_rq_clock_task() in
1571 * case they are the parent of a throttled hierarchy.
1573 if (entity_is_task(se))
1574 now = cfs_rq_clock_task(cfs_rq);
1576 now = cfs_rq_clock_task(group_cfs_rq(se));
1578 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1581 contrib_delta = __update_entity_load_avg_contrib(se);
1587 cfs_rq->runnable_load_avg += contrib_delta;
1589 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1593 * Decay the load contributed by all blocked children and account this so that
1594 * their contribution may appropriately discounted when they wake up.
1596 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1598 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1601 decays = now - cfs_rq->last_decay;
1602 if (!decays && !force_update)
1605 if (atomic_long_read(&cfs_rq->removed_load)) {
1606 unsigned long removed_load;
1607 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1608 subtract_blocked_load_contrib(cfs_rq, removed_load);
1612 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1614 atomic64_add(decays, &cfs_rq->decay_counter);
1615 cfs_rq->last_decay = now;
1618 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1621 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1623 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1624 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1627 /* Add the load generated by se into cfs_rq's child load-average */
1628 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1629 struct sched_entity *se,
1633 * We track migrations using entity decay_count <= 0, on a wake-up
1634 * migration we use a negative decay count to track the remote decays
1635 * accumulated while sleeping.
1637 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1638 * are seen by enqueue_entity_load_avg() as a migration with an already
1639 * constructed load_avg_contrib.
1641 if (unlikely(se->avg.decay_count <= 0)) {
1642 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1643 if (se->avg.decay_count) {
1645 * In a wake-up migration we have to approximate the
1646 * time sleeping. This is because we can't synchronize
1647 * clock_task between the two cpus, and it is not
1648 * guaranteed to be read-safe. Instead, we can
1649 * approximate this using our carried decays, which are
1650 * explicitly atomically readable.
1652 se->avg.last_runnable_update -= (-se->avg.decay_count)
1654 update_entity_load_avg(se, 0);
1655 /* Indicate that we're now synchronized and on-rq */
1656 se->avg.decay_count = 0;
1661 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1662 * would have made count negative); we must be careful to avoid
1663 * double-accounting blocked time after synchronizing decays.
1665 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1669 /* migrated tasks did not contribute to our blocked load */
1671 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1672 update_entity_load_avg(se, 0);
1675 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1676 /* we force update consideration on load-balancer moves */
1677 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1681 * Remove se's load from this cfs_rq child load-average, if the entity is
1682 * transitioning to a blocked state we track its projected decay using
1685 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1686 struct sched_entity *se,
1689 update_entity_load_avg(se, 1);
1690 /* we force update consideration on load-balancer moves */
1691 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1693 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1695 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1696 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1697 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1701 * Update the rq's load with the elapsed running time before entering
1702 * idle. if the last scheduled task is not a CFS task, idle_enter will
1703 * be the only way to update the runnable statistic.
1705 void idle_enter_fair(struct rq *this_rq)
1707 update_rq_runnable_avg(this_rq, 1);
1711 * Update the rq's load with the elapsed idle time before a task is
1712 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1713 * be the only way to update the runnable statistic.
1715 void idle_exit_fair(struct rq *this_rq)
1717 update_rq_runnable_avg(this_rq, 0);
1721 static inline void update_entity_load_avg(struct sched_entity *se,
1722 int update_cfs_rq) {}
1723 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1724 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1725 struct sched_entity *se,
1727 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1728 struct sched_entity *se,
1730 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1731 int force_update) {}
1734 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1736 #ifdef CONFIG_SCHEDSTATS
1737 struct task_struct *tsk = NULL;
1739 if (entity_is_task(se))
1742 if (se->statistics.sleep_start) {
1743 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1748 if (unlikely(delta > se->statistics.sleep_max))
1749 se->statistics.sleep_max = delta;
1751 se->statistics.sleep_start = 0;
1752 se->statistics.sum_sleep_runtime += delta;
1755 account_scheduler_latency(tsk, delta >> 10, 1);
1756 trace_sched_stat_sleep(tsk, delta);
1759 if (se->statistics.block_start) {
1760 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1765 if (unlikely(delta > se->statistics.block_max))
1766 se->statistics.block_max = delta;
1768 se->statistics.block_start = 0;
1769 se->statistics.sum_sleep_runtime += delta;
1772 if (tsk->in_iowait) {
1773 se->statistics.iowait_sum += delta;
1774 se->statistics.iowait_count++;
1775 trace_sched_stat_iowait(tsk, delta);
1778 trace_sched_stat_blocked(tsk, delta);
1781 * Blocking time is in units of nanosecs, so shift by
1782 * 20 to get a milliseconds-range estimation of the
1783 * amount of time that the task spent sleeping:
1785 if (unlikely(prof_on == SLEEP_PROFILING)) {
1786 profile_hits(SLEEP_PROFILING,
1787 (void *)get_wchan(tsk),
1790 account_scheduler_latency(tsk, delta >> 10, 0);
1796 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1798 #ifdef CONFIG_SCHED_DEBUG
1799 s64 d = se->vruntime - cfs_rq->min_vruntime;
1804 if (d > 3*sysctl_sched_latency)
1805 schedstat_inc(cfs_rq, nr_spread_over);
1810 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1812 u64 vruntime = cfs_rq->min_vruntime;
1815 * The 'current' period is already promised to the current tasks,
1816 * however the extra weight of the new task will slow them down a
1817 * little, place the new task so that it fits in the slot that
1818 * stays open at the end.
1820 if (initial && sched_feat(START_DEBIT))
1821 vruntime += sched_vslice(cfs_rq, se);
1823 /* sleeps up to a single latency don't count. */
1825 unsigned long thresh = sysctl_sched_latency;
1828 * Halve their sleep time's effect, to allow
1829 * for a gentler effect of sleepers:
1831 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1837 /* ensure we never gain time by being placed backwards. */
1838 se->vruntime = max_vruntime(se->vruntime, vruntime);
1841 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1844 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1847 * Update the normalized vruntime before updating min_vruntime
1848 * through calling update_curr().
1850 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1851 se->vruntime += cfs_rq->min_vruntime;
1854 * Update run-time statistics of the 'current'.
1856 update_curr(cfs_rq);
1857 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1858 account_entity_enqueue(cfs_rq, se);
1859 update_cfs_shares(cfs_rq);
1861 if (flags & ENQUEUE_WAKEUP) {
1862 place_entity(cfs_rq, se, 0);
1863 enqueue_sleeper(cfs_rq, se);
1866 update_stats_enqueue(cfs_rq, se);
1867 check_spread(cfs_rq, se);
1868 if (se != cfs_rq->curr)
1869 __enqueue_entity(cfs_rq, se);
1872 if (cfs_rq->nr_running == 1) {
1873 list_add_leaf_cfs_rq(cfs_rq);
1874 check_enqueue_throttle(cfs_rq);
1878 static void __clear_buddies_last(struct sched_entity *se)
1880 for_each_sched_entity(se) {
1881 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1882 if (cfs_rq->last == se)
1883 cfs_rq->last = NULL;
1889 static void __clear_buddies_next(struct sched_entity *se)
1891 for_each_sched_entity(se) {
1892 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1893 if (cfs_rq->next == se)
1894 cfs_rq->next = NULL;
1900 static void __clear_buddies_skip(struct sched_entity *se)
1902 for_each_sched_entity(se) {
1903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1904 if (cfs_rq->skip == se)
1905 cfs_rq->skip = NULL;
1911 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1913 if (cfs_rq->last == se)
1914 __clear_buddies_last(se);
1916 if (cfs_rq->next == se)
1917 __clear_buddies_next(se);
1919 if (cfs_rq->skip == se)
1920 __clear_buddies_skip(se);
1923 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1926 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1929 * Update run-time statistics of the 'current'.
1931 update_curr(cfs_rq);
1932 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1934 update_stats_dequeue(cfs_rq, se);
1935 if (flags & DEQUEUE_SLEEP) {
1936 #ifdef CONFIG_SCHEDSTATS
1937 if (entity_is_task(se)) {
1938 struct task_struct *tsk = task_of(se);
1940 if (tsk->state & TASK_INTERRUPTIBLE)
1941 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1942 if (tsk->state & TASK_UNINTERRUPTIBLE)
1943 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1948 clear_buddies(cfs_rq, se);
1950 if (se != cfs_rq->curr)
1951 __dequeue_entity(cfs_rq, se);
1953 account_entity_dequeue(cfs_rq, se);
1956 * Normalize the entity after updating the min_vruntime because the
1957 * update can refer to the ->curr item and we need to reflect this
1958 * movement in our normalized position.
1960 if (!(flags & DEQUEUE_SLEEP))
1961 se->vruntime -= cfs_rq->min_vruntime;
1963 /* return excess runtime on last dequeue */
1964 return_cfs_rq_runtime(cfs_rq);
1966 update_min_vruntime(cfs_rq);
1967 update_cfs_shares(cfs_rq);
1971 * Preempt the current task with a newly woken task if needed:
1974 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1976 unsigned long ideal_runtime, delta_exec;
1977 struct sched_entity *se;
1980 ideal_runtime = sched_slice(cfs_rq, curr);
1981 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1982 if (delta_exec > ideal_runtime) {
1983 resched_task(rq_of(cfs_rq)->curr);
1985 * The current task ran long enough, ensure it doesn't get
1986 * re-elected due to buddy favours.
1988 clear_buddies(cfs_rq, curr);
1993 * Ensure that a task that missed wakeup preemption by a
1994 * narrow margin doesn't have to wait for a full slice.
1995 * This also mitigates buddy induced latencies under load.
1997 if (delta_exec < sysctl_sched_min_granularity)
2000 se = __pick_first_entity(cfs_rq);
2001 delta = curr->vruntime - se->vruntime;
2006 if (delta > ideal_runtime)
2007 resched_task(rq_of(cfs_rq)->curr);
2011 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2013 /* 'current' is not kept within the tree. */
2016 * Any task has to be enqueued before it get to execute on
2017 * a CPU. So account for the time it spent waiting on the
2020 update_stats_wait_end(cfs_rq, se);
2021 __dequeue_entity(cfs_rq, se);
2024 update_stats_curr_start(cfs_rq, se);
2026 #ifdef CONFIG_SCHEDSTATS
2028 * Track our maximum slice length, if the CPU's load is at
2029 * least twice that of our own weight (i.e. dont track it
2030 * when there are only lesser-weight tasks around):
2032 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2033 se->statistics.slice_max = max(se->statistics.slice_max,
2034 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2037 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2041 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2044 * Pick the next process, keeping these things in mind, in this order:
2045 * 1) keep things fair between processes/task groups
2046 * 2) pick the "next" process, since someone really wants that to run
2047 * 3) pick the "last" process, for cache locality
2048 * 4) do not run the "skip" process, if something else is available
2050 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2052 struct sched_entity *se = __pick_first_entity(cfs_rq);
2053 struct sched_entity *left = se;
2056 * Avoid running the skip buddy, if running something else can
2057 * be done without getting too unfair.
2059 if (cfs_rq->skip == se) {
2060 struct sched_entity *second = __pick_next_entity(se);
2061 if (second && wakeup_preempt_entity(second, left) < 1)
2066 * Prefer last buddy, try to return the CPU to a preempted task.
2068 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2072 * Someone really wants this to run. If it's not unfair, run it.
2074 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2077 clear_buddies(cfs_rq, se);
2082 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2084 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2087 * If still on the runqueue then deactivate_task()
2088 * was not called and update_curr() has to be done:
2091 update_curr(cfs_rq);
2093 /* throttle cfs_rqs exceeding runtime */
2094 check_cfs_rq_runtime(cfs_rq);
2096 check_spread(cfs_rq, prev);
2098 update_stats_wait_start(cfs_rq, prev);
2099 /* Put 'current' back into the tree. */
2100 __enqueue_entity(cfs_rq, prev);
2101 /* in !on_rq case, update occurred at dequeue */
2102 update_entity_load_avg(prev, 1);
2104 cfs_rq->curr = NULL;
2108 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2111 * Update run-time statistics of the 'current'.
2113 update_curr(cfs_rq);
2116 * Ensure that runnable average is periodically updated.
2118 update_entity_load_avg(curr, 1);
2119 update_cfs_rq_blocked_load(cfs_rq, 1);
2120 update_cfs_shares(cfs_rq);
2122 #ifdef CONFIG_SCHED_HRTICK
2124 * queued ticks are scheduled to match the slice, so don't bother
2125 * validating it and just reschedule.
2128 resched_task(rq_of(cfs_rq)->curr);
2132 * don't let the period tick interfere with the hrtick preemption
2134 if (!sched_feat(DOUBLE_TICK) &&
2135 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2139 if (cfs_rq->nr_running > 1)
2140 check_preempt_tick(cfs_rq, curr);
2144 /**************************************************
2145 * CFS bandwidth control machinery
2148 #ifdef CONFIG_CFS_BANDWIDTH
2150 #ifdef HAVE_JUMP_LABEL
2151 static struct static_key __cfs_bandwidth_used;
2153 static inline bool cfs_bandwidth_used(void)
2155 return static_key_false(&__cfs_bandwidth_used);
2158 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2160 /* only need to count groups transitioning between enabled/!enabled */
2161 if (enabled && !was_enabled)
2162 static_key_slow_inc(&__cfs_bandwidth_used);
2163 else if (!enabled && was_enabled)
2164 static_key_slow_dec(&__cfs_bandwidth_used);
2166 #else /* HAVE_JUMP_LABEL */
2167 static bool cfs_bandwidth_used(void)
2172 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2173 #endif /* HAVE_JUMP_LABEL */
2176 * default period for cfs group bandwidth.
2177 * default: 0.1s, units: nanoseconds
2179 static inline u64 default_cfs_period(void)
2181 return 100000000ULL;
2184 static inline u64 sched_cfs_bandwidth_slice(void)
2186 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2190 * Replenish runtime according to assigned quota and update expiration time.
2191 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2192 * additional synchronization around rq->lock.
2194 * requires cfs_b->lock
2196 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2200 if (cfs_b->quota == RUNTIME_INF)
2203 now = sched_clock_cpu(smp_processor_id());
2204 cfs_b->runtime = cfs_b->quota;
2205 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2208 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2210 return &tg->cfs_bandwidth;
2213 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2214 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2216 if (unlikely(cfs_rq->throttle_count))
2217 return cfs_rq->throttled_clock_task;
2219 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2222 /* returns 0 on failure to allocate runtime */
2223 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2225 struct task_group *tg = cfs_rq->tg;
2226 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2227 u64 amount = 0, min_amount, expires;
2229 /* note: this is a positive sum as runtime_remaining <= 0 */
2230 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2232 raw_spin_lock(&cfs_b->lock);
2233 if (cfs_b->quota == RUNTIME_INF)
2234 amount = min_amount;
2237 * If the bandwidth pool has become inactive, then at least one
2238 * period must have elapsed since the last consumption.
2239 * Refresh the global state and ensure bandwidth timer becomes
2242 if (!cfs_b->timer_active) {
2243 __refill_cfs_bandwidth_runtime(cfs_b);
2244 __start_cfs_bandwidth(cfs_b);
2247 if (cfs_b->runtime > 0) {
2248 amount = min(cfs_b->runtime, min_amount);
2249 cfs_b->runtime -= amount;
2253 expires = cfs_b->runtime_expires;
2254 raw_spin_unlock(&cfs_b->lock);
2256 cfs_rq->runtime_remaining += amount;
2258 * we may have advanced our local expiration to account for allowed
2259 * spread between our sched_clock and the one on which runtime was
2262 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2263 cfs_rq->runtime_expires = expires;
2265 return cfs_rq->runtime_remaining > 0;
2269 * Note: This depends on the synchronization provided by sched_clock and the
2270 * fact that rq->clock snapshots this value.
2272 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2274 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2276 /* if the deadline is ahead of our clock, nothing to do */
2277 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2280 if (cfs_rq->runtime_remaining < 0)
2284 * If the local deadline has passed we have to consider the
2285 * possibility that our sched_clock is 'fast' and the global deadline
2286 * has not truly expired.
2288 * Fortunately we can check determine whether this the case by checking
2289 * whether the global deadline has advanced.
2292 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2293 /* extend local deadline, drift is bounded above by 2 ticks */
2294 cfs_rq->runtime_expires += TICK_NSEC;
2296 /* global deadline is ahead, expiration has passed */
2297 cfs_rq->runtime_remaining = 0;
2301 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2302 unsigned long delta_exec)
2304 /* dock delta_exec before expiring quota (as it could span periods) */
2305 cfs_rq->runtime_remaining -= delta_exec;
2306 expire_cfs_rq_runtime(cfs_rq);
2308 if (likely(cfs_rq->runtime_remaining > 0))
2312 * if we're unable to extend our runtime we resched so that the active
2313 * hierarchy can be throttled
2315 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2316 resched_task(rq_of(cfs_rq)->curr);
2319 static __always_inline
2320 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2322 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2325 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2328 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2330 return cfs_bandwidth_used() && cfs_rq->throttled;
2333 /* check whether cfs_rq, or any parent, is throttled */
2334 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2336 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2340 * Ensure that neither of the group entities corresponding to src_cpu or
2341 * dest_cpu are members of a throttled hierarchy when performing group
2342 * load-balance operations.
2344 static inline int throttled_lb_pair(struct task_group *tg,
2345 int src_cpu, int dest_cpu)
2347 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2349 src_cfs_rq = tg->cfs_rq[src_cpu];
2350 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2352 return throttled_hierarchy(src_cfs_rq) ||
2353 throttled_hierarchy(dest_cfs_rq);
2356 /* updated child weight may affect parent so we have to do this bottom up */
2357 static int tg_unthrottle_up(struct task_group *tg, void *data)
2359 struct rq *rq = data;
2360 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2362 cfs_rq->throttle_count--;
2364 if (!cfs_rq->throttle_count) {
2365 /* adjust cfs_rq_clock_task() */
2366 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2367 cfs_rq->throttled_clock_task;
2374 static int tg_throttle_down(struct task_group *tg, void *data)
2376 struct rq *rq = data;
2377 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2379 /* group is entering throttled state, stop time */
2380 if (!cfs_rq->throttle_count)
2381 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2382 cfs_rq->throttle_count++;
2387 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2389 struct rq *rq = rq_of(cfs_rq);
2390 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2391 struct sched_entity *se;
2392 long task_delta, dequeue = 1;
2394 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2396 /* freeze hierarchy runnable averages while throttled */
2398 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2401 task_delta = cfs_rq->h_nr_running;
2402 for_each_sched_entity(se) {
2403 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2404 /* throttled entity or throttle-on-deactivate */
2409 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2410 qcfs_rq->h_nr_running -= task_delta;
2412 if (qcfs_rq->load.weight)
2417 rq->nr_running -= task_delta;
2419 cfs_rq->throttled = 1;
2420 cfs_rq->throttled_clock = rq_clock(rq);
2421 raw_spin_lock(&cfs_b->lock);
2422 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2423 raw_spin_unlock(&cfs_b->lock);
2426 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2428 struct rq *rq = rq_of(cfs_rq);
2429 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2430 struct sched_entity *se;
2434 se = cfs_rq->tg->se[cpu_of(rq)];
2436 cfs_rq->throttled = 0;
2438 update_rq_clock(rq);
2440 raw_spin_lock(&cfs_b->lock);
2441 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2442 list_del_rcu(&cfs_rq->throttled_list);
2443 raw_spin_unlock(&cfs_b->lock);
2445 /* update hierarchical throttle state */
2446 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2448 if (!cfs_rq->load.weight)
2451 task_delta = cfs_rq->h_nr_running;
2452 for_each_sched_entity(se) {
2456 cfs_rq = cfs_rq_of(se);
2458 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2459 cfs_rq->h_nr_running += task_delta;
2461 if (cfs_rq_throttled(cfs_rq))
2466 rq->nr_running += task_delta;
2468 /* determine whether we need to wake up potentially idle cpu */
2469 if (rq->curr == rq->idle && rq->cfs.nr_running)
2470 resched_task(rq->curr);
2473 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2474 u64 remaining, u64 expires)
2476 struct cfs_rq *cfs_rq;
2477 u64 runtime = remaining;
2480 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2482 struct rq *rq = rq_of(cfs_rq);
2484 raw_spin_lock(&rq->lock);
2485 if (!cfs_rq_throttled(cfs_rq))
2488 runtime = -cfs_rq->runtime_remaining + 1;
2489 if (runtime > remaining)
2490 runtime = remaining;
2491 remaining -= runtime;
2493 cfs_rq->runtime_remaining += runtime;
2494 cfs_rq->runtime_expires = expires;
2496 /* we check whether we're throttled above */
2497 if (cfs_rq->runtime_remaining > 0)
2498 unthrottle_cfs_rq(cfs_rq);
2501 raw_spin_unlock(&rq->lock);
2512 * Responsible for refilling a task_group's bandwidth and unthrottling its
2513 * cfs_rqs as appropriate. If there has been no activity within the last
2514 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2515 * used to track this state.
2517 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2519 u64 runtime, runtime_expires;
2520 int idle = 1, throttled;
2522 raw_spin_lock(&cfs_b->lock);
2523 /* no need to continue the timer with no bandwidth constraint */
2524 if (cfs_b->quota == RUNTIME_INF)
2527 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2528 /* idle depends on !throttled (for the case of a large deficit) */
2529 idle = cfs_b->idle && !throttled;
2530 cfs_b->nr_periods += overrun;
2532 /* if we're going inactive then everything else can be deferred */
2536 __refill_cfs_bandwidth_runtime(cfs_b);
2539 /* mark as potentially idle for the upcoming period */
2544 /* account preceding periods in which throttling occurred */
2545 cfs_b->nr_throttled += overrun;
2548 * There are throttled entities so we must first use the new bandwidth
2549 * to unthrottle them before making it generally available. This
2550 * ensures that all existing debts will be paid before a new cfs_rq is
2553 runtime = cfs_b->runtime;
2554 runtime_expires = cfs_b->runtime_expires;
2558 * This check is repeated as we are holding onto the new bandwidth
2559 * while we unthrottle. This can potentially race with an unthrottled
2560 * group trying to acquire new bandwidth from the global pool.
2562 while (throttled && runtime > 0) {
2563 raw_spin_unlock(&cfs_b->lock);
2564 /* we can't nest cfs_b->lock while distributing bandwidth */
2565 runtime = distribute_cfs_runtime(cfs_b, runtime,
2567 raw_spin_lock(&cfs_b->lock);
2569 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2572 /* return (any) remaining runtime */
2573 cfs_b->runtime = runtime;
2575 * While we are ensured activity in the period following an
2576 * unthrottle, this also covers the case in which the new bandwidth is
2577 * insufficient to cover the existing bandwidth deficit. (Forcing the
2578 * timer to remain active while there are any throttled entities.)
2583 cfs_b->timer_active = 0;
2584 raw_spin_unlock(&cfs_b->lock);
2589 /* a cfs_rq won't donate quota below this amount */
2590 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2591 /* minimum remaining period time to redistribute slack quota */
2592 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2593 /* how long we wait to gather additional slack before distributing */
2594 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2596 /* are we near the end of the current quota period? */
2597 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2599 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2602 /* if the call-back is running a quota refresh is already occurring */
2603 if (hrtimer_callback_running(refresh_timer))
2606 /* is a quota refresh about to occur? */
2607 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2608 if (remaining < min_expire)
2614 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2616 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2618 /* if there's a quota refresh soon don't bother with slack */
2619 if (runtime_refresh_within(cfs_b, min_left))
2622 start_bandwidth_timer(&cfs_b->slack_timer,
2623 ns_to_ktime(cfs_bandwidth_slack_period));
2626 /* we know any runtime found here is valid as update_curr() precedes return */
2627 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2629 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2630 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2632 if (slack_runtime <= 0)
2635 raw_spin_lock(&cfs_b->lock);
2636 if (cfs_b->quota != RUNTIME_INF &&
2637 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2638 cfs_b->runtime += slack_runtime;
2640 /* we are under rq->lock, defer unthrottling using a timer */
2641 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2642 !list_empty(&cfs_b->throttled_cfs_rq))
2643 start_cfs_slack_bandwidth(cfs_b);
2645 raw_spin_unlock(&cfs_b->lock);
2647 /* even if it's not valid for return we don't want to try again */
2648 cfs_rq->runtime_remaining -= slack_runtime;
2651 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2653 if (!cfs_bandwidth_used())
2656 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2659 __return_cfs_rq_runtime(cfs_rq);
2663 * This is done with a timer (instead of inline with bandwidth return) since
2664 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2666 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2668 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2671 /* confirm we're still not at a refresh boundary */
2672 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2675 raw_spin_lock(&cfs_b->lock);
2676 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2677 runtime = cfs_b->runtime;
2680 expires = cfs_b->runtime_expires;
2681 raw_spin_unlock(&cfs_b->lock);
2686 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2688 raw_spin_lock(&cfs_b->lock);
2689 if (expires == cfs_b->runtime_expires)
2690 cfs_b->runtime = runtime;
2691 raw_spin_unlock(&cfs_b->lock);
2695 * When a group wakes up we want to make sure that its quota is not already
2696 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2697 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2699 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2701 if (!cfs_bandwidth_used())
2704 /* an active group must be handled by the update_curr()->put() path */
2705 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2708 /* ensure the group is not already throttled */
2709 if (cfs_rq_throttled(cfs_rq))
2712 /* update runtime allocation */
2713 account_cfs_rq_runtime(cfs_rq, 0);
2714 if (cfs_rq->runtime_remaining <= 0)
2715 throttle_cfs_rq(cfs_rq);
2718 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2719 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2721 if (!cfs_bandwidth_used())
2724 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2728 * it's possible for a throttled entity to be forced into a running
2729 * state (e.g. set_curr_task), in this case we're finished.
2731 if (cfs_rq_throttled(cfs_rq))
2734 throttle_cfs_rq(cfs_rq);
2737 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2739 struct cfs_bandwidth *cfs_b =
2740 container_of(timer, struct cfs_bandwidth, slack_timer);
2741 do_sched_cfs_slack_timer(cfs_b);
2743 return HRTIMER_NORESTART;
2746 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2748 struct cfs_bandwidth *cfs_b =
2749 container_of(timer, struct cfs_bandwidth, period_timer);
2755 now = hrtimer_cb_get_time(timer);
2756 overrun = hrtimer_forward(timer, now, cfs_b->period);
2761 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2764 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2767 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2769 raw_spin_lock_init(&cfs_b->lock);
2771 cfs_b->quota = RUNTIME_INF;
2772 cfs_b->period = ns_to_ktime(default_cfs_period());
2774 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2775 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2776 cfs_b->period_timer.function = sched_cfs_period_timer;
2777 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2778 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2781 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2783 cfs_rq->runtime_enabled = 0;
2784 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2787 /* requires cfs_b->lock, may release to reprogram timer */
2788 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2791 * The timer may be active because we're trying to set a new bandwidth
2792 * period or because we're racing with the tear-down path
2793 * (timer_active==0 becomes visible before the hrtimer call-back
2794 * terminates). In either case we ensure that it's re-programmed
2796 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2797 raw_spin_unlock(&cfs_b->lock);
2798 /* ensure cfs_b->lock is available while we wait */
2799 hrtimer_cancel(&cfs_b->period_timer);
2801 raw_spin_lock(&cfs_b->lock);
2802 /* if someone else restarted the timer then we're done */
2803 if (cfs_b->timer_active)
2807 cfs_b->timer_active = 1;
2808 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2811 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2813 hrtimer_cancel(&cfs_b->period_timer);
2814 hrtimer_cancel(&cfs_b->slack_timer);
2817 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2819 struct cfs_rq *cfs_rq;
2821 for_each_leaf_cfs_rq(rq, cfs_rq) {
2822 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2824 if (!cfs_rq->runtime_enabled)
2828 * clock_task is not advancing so we just need to make sure
2829 * there's some valid quota amount
2831 cfs_rq->runtime_remaining = cfs_b->quota;
2832 if (cfs_rq_throttled(cfs_rq))
2833 unthrottle_cfs_rq(cfs_rq);
2837 #else /* CONFIG_CFS_BANDWIDTH */
2838 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2840 return rq_clock_task(rq_of(cfs_rq));
2843 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2844 unsigned long delta_exec) {}
2845 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2846 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2847 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2849 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2854 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2859 static inline int throttled_lb_pair(struct task_group *tg,
2860 int src_cpu, int dest_cpu)
2865 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2867 #ifdef CONFIG_FAIR_GROUP_SCHED
2868 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2871 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2875 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2876 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2878 #endif /* CONFIG_CFS_BANDWIDTH */
2880 /**************************************************
2881 * CFS operations on tasks:
2884 #ifdef CONFIG_SCHED_HRTICK
2885 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2887 struct sched_entity *se = &p->se;
2888 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2890 WARN_ON(task_rq(p) != rq);
2892 if (cfs_rq->nr_running > 1) {
2893 u64 slice = sched_slice(cfs_rq, se);
2894 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2895 s64 delta = slice - ran;
2904 * Don't schedule slices shorter than 10000ns, that just
2905 * doesn't make sense. Rely on vruntime for fairness.
2908 delta = max_t(s64, 10000LL, delta);
2910 hrtick_start(rq, delta);
2915 * called from enqueue/dequeue and updates the hrtick when the
2916 * current task is from our class and nr_running is low enough
2919 static void hrtick_update(struct rq *rq)
2921 struct task_struct *curr = rq->curr;
2923 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2926 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2927 hrtick_start_fair(rq, curr);
2929 #else /* !CONFIG_SCHED_HRTICK */
2931 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2935 static inline void hrtick_update(struct rq *rq)
2941 * The enqueue_task method is called before nr_running is
2942 * increased. Here we update the fair scheduling stats and
2943 * then put the task into the rbtree:
2946 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2948 struct cfs_rq *cfs_rq;
2949 struct sched_entity *se = &p->se;
2951 for_each_sched_entity(se) {
2954 cfs_rq = cfs_rq_of(se);
2955 enqueue_entity(cfs_rq, se, flags);
2958 * end evaluation on encountering a throttled cfs_rq
2960 * note: in the case of encountering a throttled cfs_rq we will
2961 * post the final h_nr_running increment below.
2963 if (cfs_rq_throttled(cfs_rq))
2965 cfs_rq->h_nr_running++;
2967 flags = ENQUEUE_WAKEUP;
2970 for_each_sched_entity(se) {
2971 cfs_rq = cfs_rq_of(se);
2972 cfs_rq->h_nr_running++;
2974 if (cfs_rq_throttled(cfs_rq))
2977 update_cfs_shares(cfs_rq);
2978 update_entity_load_avg(se, 1);
2982 update_rq_runnable_avg(rq, rq->nr_running);
2988 static void set_next_buddy(struct sched_entity *se);
2991 * The dequeue_task method is called before nr_running is
2992 * decreased. We remove the task from the rbtree and
2993 * update the fair scheduling stats:
2995 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2997 struct cfs_rq *cfs_rq;
2998 struct sched_entity *se = &p->se;
2999 int task_sleep = flags & DEQUEUE_SLEEP;
3001 for_each_sched_entity(se) {
3002 cfs_rq = cfs_rq_of(se);
3003 dequeue_entity(cfs_rq, se, flags);
3006 * end evaluation on encountering a throttled cfs_rq
3008 * note: in the case of encountering a throttled cfs_rq we will
3009 * post the final h_nr_running decrement below.
3011 if (cfs_rq_throttled(cfs_rq))
3013 cfs_rq->h_nr_running--;
3015 /* Don't dequeue parent if it has other entities besides us */
3016 if (cfs_rq->load.weight) {
3018 * Bias pick_next to pick a task from this cfs_rq, as
3019 * p is sleeping when it is within its sched_slice.
3021 if (task_sleep && parent_entity(se))
3022 set_next_buddy(parent_entity(se));
3024 /* avoid re-evaluating load for this entity */
3025 se = parent_entity(se);
3028 flags |= DEQUEUE_SLEEP;
3031 for_each_sched_entity(se) {
3032 cfs_rq = cfs_rq_of(se);
3033 cfs_rq->h_nr_running--;
3035 if (cfs_rq_throttled(cfs_rq))
3038 update_cfs_shares(cfs_rq);
3039 update_entity_load_avg(se, 1);
3044 update_rq_runnable_avg(rq, 1);
3050 /* Used instead of source_load when we know the type == 0 */
3051 static unsigned long weighted_cpuload(const int cpu)
3053 return cpu_rq(cpu)->cfs.runnable_load_avg;
3057 * Return a low guess at the load of a migration-source cpu weighted
3058 * according to the scheduling class and "nice" value.
3060 * We want to under-estimate the load of migration sources, to
3061 * balance conservatively.
3063 static unsigned long source_load(int cpu, int type)
3065 struct rq *rq = cpu_rq(cpu);
3066 unsigned long total = weighted_cpuload(cpu);
3068 if (type == 0 || !sched_feat(LB_BIAS))
3071 return min(rq->cpu_load[type-1], total);
3075 * Return a high guess at the load of a migration-target cpu weighted
3076 * according to the scheduling class and "nice" value.
3078 static unsigned long target_load(int cpu, int type)
3080 struct rq *rq = cpu_rq(cpu);
3081 unsigned long total = weighted_cpuload(cpu);
3083 if (type == 0 || !sched_feat(LB_BIAS))
3086 return max(rq->cpu_load[type-1], total);
3089 static unsigned long power_of(int cpu)
3091 return cpu_rq(cpu)->cpu_power;
3094 static unsigned long cpu_avg_load_per_task(int cpu)
3096 struct rq *rq = cpu_rq(cpu);
3097 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3098 unsigned long load_avg = rq->cfs.runnable_load_avg;
3101 return load_avg / nr_running;
3106 static void record_wakee(struct task_struct *p)
3109 * Rough decay (wiping) for cost saving, don't worry
3110 * about the boundary, really active task won't care
3113 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3114 current->wakee_flips = 0;
3115 current->wakee_flip_decay_ts = jiffies;
3118 if (current->last_wakee != p) {
3119 current->last_wakee = p;
3120 current->wakee_flips++;
3124 static void task_waking_fair(struct task_struct *p)
3126 struct sched_entity *se = &p->se;
3127 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3130 #ifndef CONFIG_64BIT
3131 u64 min_vruntime_copy;
3134 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3136 min_vruntime = cfs_rq->min_vruntime;
3137 } while (min_vruntime != min_vruntime_copy);
3139 min_vruntime = cfs_rq->min_vruntime;
3142 se->vruntime -= min_vruntime;
3146 #ifdef CONFIG_FAIR_GROUP_SCHED
3148 * effective_load() calculates the load change as seen from the root_task_group
3150 * Adding load to a group doesn't make a group heavier, but can cause movement
3151 * of group shares between cpus. Assuming the shares were perfectly aligned one
3152 * can calculate the shift in shares.
3154 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3155 * on this @cpu and results in a total addition (subtraction) of @wg to the
3156 * total group weight.
3158 * Given a runqueue weight distribution (rw_i) we can compute a shares
3159 * distribution (s_i) using:
3161 * s_i = rw_i / \Sum rw_j (1)
3163 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3164 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3165 * shares distribution (s_i):
3167 * rw_i = { 2, 4, 1, 0 }
3168 * s_i = { 2/7, 4/7, 1/7, 0 }
3170 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3171 * task used to run on and the CPU the waker is running on), we need to
3172 * compute the effect of waking a task on either CPU and, in case of a sync
3173 * wakeup, compute the effect of the current task going to sleep.
3175 * So for a change of @wl to the local @cpu with an overall group weight change
3176 * of @wl we can compute the new shares distribution (s'_i) using:
3178 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3180 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3181 * differences in waking a task to CPU 0. The additional task changes the
3182 * weight and shares distributions like:
3184 * rw'_i = { 3, 4, 1, 0 }
3185 * s'_i = { 3/8, 4/8, 1/8, 0 }
3187 * We can then compute the difference in effective weight by using:
3189 * dw_i = S * (s'_i - s_i) (3)
3191 * Where 'S' is the group weight as seen by its parent.
3193 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3194 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3195 * 4/7) times the weight of the group.
3197 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3199 struct sched_entity *se = tg->se[cpu];
3201 if (!tg->parent) /* the trivial, non-cgroup case */
3204 for_each_sched_entity(se) {
3210 * W = @wg + \Sum rw_j
3212 W = wg + calc_tg_weight(tg, se->my_q);
3217 w = se->my_q->load.weight + wl;
3220 * wl = S * s'_i; see (2)
3223 wl = (w * tg->shares) / W;
3228 * Per the above, wl is the new se->load.weight value; since
3229 * those are clipped to [MIN_SHARES, ...) do so now. See
3230 * calc_cfs_shares().
3232 if (wl < MIN_SHARES)
3236 * wl = dw_i = S * (s'_i - s_i); see (3)
3238 wl -= se->load.weight;
3241 * Recursively apply this logic to all parent groups to compute
3242 * the final effective load change on the root group. Since
3243 * only the @tg group gets extra weight, all parent groups can
3244 * only redistribute existing shares. @wl is the shift in shares
3245 * resulting from this level per the above.
3254 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3255 unsigned long wl, unsigned long wg)
3262 static int wake_wide(struct task_struct *p)
3264 int factor = this_cpu_read(sd_llc_size);
3267 * Yeah, it's the switching-frequency, could means many wakee or
3268 * rapidly switch, use factor here will just help to automatically
3269 * adjust the loose-degree, so bigger node will lead to more pull.
3271 if (p->wakee_flips > factor) {
3273 * wakee is somewhat hot, it needs certain amount of cpu
3274 * resource, so if waker is far more hot, prefer to leave
3277 if (current->wakee_flips > (factor * p->wakee_flips))
3284 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3286 s64 this_load, load;
3287 int idx, this_cpu, prev_cpu;
3288 unsigned long tl_per_task;
3289 struct task_group *tg;
3290 unsigned long weight;
3294 * If we wake multiple tasks be careful to not bounce
3295 * ourselves around too much.
3301 this_cpu = smp_processor_id();
3302 prev_cpu = task_cpu(p);
3303 load = source_load(prev_cpu, idx);
3304 this_load = target_load(this_cpu, idx);
3307 * If sync wakeup then subtract the (maximum possible)
3308 * effect of the currently running task from the load
3309 * of the current CPU:
3312 tg = task_group(current);
3313 weight = current->se.load.weight;
3315 this_load += effective_load(tg, this_cpu, -weight, -weight);
3316 load += effective_load(tg, prev_cpu, 0, -weight);
3320 weight = p->se.load.weight;
3323 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3324 * due to the sync cause above having dropped this_load to 0, we'll
3325 * always have an imbalance, but there's really nothing you can do
3326 * about that, so that's good too.
3328 * Otherwise check if either cpus are near enough in load to allow this
3329 * task to be woken on this_cpu.
3331 if (this_load > 0) {
3332 s64 this_eff_load, prev_eff_load;
3334 this_eff_load = 100;
3335 this_eff_load *= power_of(prev_cpu);
3336 this_eff_load *= this_load +
3337 effective_load(tg, this_cpu, weight, weight);
3339 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3340 prev_eff_load *= power_of(this_cpu);
3341 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3343 balanced = this_eff_load <= prev_eff_load;
3348 * If the currently running task will sleep within
3349 * a reasonable amount of time then attract this newly
3352 if (sync && balanced)
3355 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3356 tl_per_task = cpu_avg_load_per_task(this_cpu);
3359 (this_load <= load &&
3360 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3362 * This domain has SD_WAKE_AFFINE and
3363 * p is cache cold in this domain, and
3364 * there is no bad imbalance.
3366 schedstat_inc(sd, ttwu_move_affine);
3367 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3375 * find_idlest_group finds and returns the least busy CPU group within the
3378 static struct sched_group *
3379 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3380 int this_cpu, int load_idx)
3382 struct sched_group *idlest = NULL, *group = sd->groups;
3383 unsigned long min_load = ULONG_MAX, this_load = 0;
3384 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3387 unsigned long load, avg_load;
3391 /* Skip over this group if it has no CPUs allowed */
3392 if (!cpumask_intersects(sched_group_cpus(group),
3393 tsk_cpus_allowed(p)))
3396 local_group = cpumask_test_cpu(this_cpu,
3397 sched_group_cpus(group));
3399 /* Tally up the load of all CPUs in the group */
3402 for_each_cpu(i, sched_group_cpus(group)) {
3403 /* Bias balancing toward cpus of our domain */
3405 load = source_load(i, load_idx);
3407 load = target_load(i, load_idx);
3412 /* Adjust by relative CPU power of the group */
3413 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3416 this_load = avg_load;
3417 } else if (avg_load < min_load) {
3418 min_load = avg_load;
3421 } while (group = group->next, group != sd->groups);
3423 if (!idlest || 100*this_load < imbalance*min_load)
3429 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3432 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3434 unsigned long load, min_load = ULONG_MAX;
3438 /* Traverse only the allowed CPUs */
3439 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3440 load = weighted_cpuload(i);
3442 if (load < min_load || (load == min_load && i == this_cpu)) {
3452 * Try and locate an idle CPU in the sched_domain.
3454 static int select_idle_sibling(struct task_struct *p, int target)
3456 struct sched_domain *sd;
3457 struct sched_group *sg;
3458 int i = task_cpu(p);
3460 if (idle_cpu(target))
3464 * If the prevous cpu is cache affine and idle, don't be stupid.
3466 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3470 * Otherwise, iterate the domains and find an elegible idle cpu.
3472 sd = rcu_dereference(per_cpu(sd_llc, target));
3473 for_each_lower_domain(sd) {
3476 if (!cpumask_intersects(sched_group_cpus(sg),
3477 tsk_cpus_allowed(p)))
3480 for_each_cpu(i, sched_group_cpus(sg)) {
3481 if (i == target || !idle_cpu(i))
3485 target = cpumask_first_and(sched_group_cpus(sg),
3486 tsk_cpus_allowed(p));
3490 } while (sg != sd->groups);
3497 * sched_balance_self: balance the current task (running on cpu) in domains
3498 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3501 * Balance, ie. select the least loaded group.
3503 * Returns the target CPU number, or the same CPU if no balancing is needed.
3505 * preempt must be disabled.
3508 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3510 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3511 int cpu = smp_processor_id();
3512 int prev_cpu = task_cpu(p);
3514 int want_affine = 0;
3515 int sync = wake_flags & WF_SYNC;
3517 if (p->nr_cpus_allowed == 1)
3520 if (sd_flag & SD_BALANCE_WAKE) {
3521 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3527 for_each_domain(cpu, tmp) {
3528 if (!(tmp->flags & SD_LOAD_BALANCE))
3532 * If both cpu and prev_cpu are part of this domain,
3533 * cpu is a valid SD_WAKE_AFFINE target.
3535 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3536 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3541 if (tmp->flags & sd_flag)
3546 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3549 new_cpu = select_idle_sibling(p, prev_cpu);
3554 int load_idx = sd->forkexec_idx;
3555 struct sched_group *group;
3558 if (!(sd->flags & sd_flag)) {
3563 if (sd_flag & SD_BALANCE_WAKE)
3564 load_idx = sd->wake_idx;
3566 group = find_idlest_group(sd, p, cpu, load_idx);
3572 new_cpu = find_idlest_cpu(group, p, cpu);
3573 if (new_cpu == -1 || new_cpu == cpu) {
3574 /* Now try balancing at a lower domain level of cpu */
3579 /* Now try balancing at a lower domain level of new_cpu */
3581 weight = sd->span_weight;
3583 for_each_domain(cpu, tmp) {
3584 if (weight <= tmp->span_weight)
3586 if (tmp->flags & sd_flag)
3589 /* while loop will break here if sd == NULL */
3598 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3599 * cfs_rq_of(p) references at time of call are still valid and identify the
3600 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3601 * other assumptions, including the state of rq->lock, should be made.
3604 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3606 struct sched_entity *se = &p->se;
3607 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3610 * Load tracking: accumulate removed load so that it can be processed
3611 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3612 * to blocked load iff they have a positive decay-count. It can never
3613 * be negative here since on-rq tasks have decay-count == 0.
3615 if (se->avg.decay_count) {
3616 se->avg.decay_count = -__synchronize_entity_decay(se);
3617 atomic_long_add(se->avg.load_avg_contrib,
3618 &cfs_rq->removed_load);
3621 #endif /* CONFIG_SMP */
3623 static unsigned long
3624 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3626 unsigned long gran = sysctl_sched_wakeup_granularity;
3629 * Since its curr running now, convert the gran from real-time
3630 * to virtual-time in his units.
3632 * By using 'se' instead of 'curr' we penalize light tasks, so
3633 * they get preempted easier. That is, if 'se' < 'curr' then
3634 * the resulting gran will be larger, therefore penalizing the
3635 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3636 * be smaller, again penalizing the lighter task.
3638 * This is especially important for buddies when the leftmost
3639 * task is higher priority than the buddy.
3641 return calc_delta_fair(gran, se);
3645 * Should 'se' preempt 'curr'.
3659 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3661 s64 gran, vdiff = curr->vruntime - se->vruntime;
3666 gran = wakeup_gran(curr, se);
3673 static void set_last_buddy(struct sched_entity *se)
3675 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3678 for_each_sched_entity(se)
3679 cfs_rq_of(se)->last = se;
3682 static void set_next_buddy(struct sched_entity *se)
3684 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3687 for_each_sched_entity(se)
3688 cfs_rq_of(se)->next = se;
3691 static void set_skip_buddy(struct sched_entity *se)
3693 for_each_sched_entity(se)
3694 cfs_rq_of(se)->skip = se;
3698 * Preempt the current task with a newly woken task if needed:
3700 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3702 struct task_struct *curr = rq->curr;
3703 struct sched_entity *se = &curr->se, *pse = &p->se;
3704 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3705 int scale = cfs_rq->nr_running >= sched_nr_latency;
3706 int next_buddy_marked = 0;
3708 if (unlikely(se == pse))
3712 * This is possible from callers such as move_task(), in which we
3713 * unconditionally check_prempt_curr() after an enqueue (which may have
3714 * lead to a throttle). This both saves work and prevents false
3715 * next-buddy nomination below.
3717 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3720 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3721 set_next_buddy(pse);
3722 next_buddy_marked = 1;
3726 * We can come here with TIF_NEED_RESCHED already set from new task
3729 * Note: this also catches the edge-case of curr being in a throttled
3730 * group (e.g. via set_curr_task), since update_curr() (in the
3731 * enqueue of curr) will have resulted in resched being set. This
3732 * prevents us from potentially nominating it as a false LAST_BUDDY
3735 if (test_tsk_need_resched(curr))
3738 /* Idle tasks are by definition preempted by non-idle tasks. */
3739 if (unlikely(curr->policy == SCHED_IDLE) &&
3740 likely(p->policy != SCHED_IDLE))
3744 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3745 * is driven by the tick):
3747 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3750 find_matching_se(&se, &pse);
3751 update_curr(cfs_rq_of(se));
3753 if (wakeup_preempt_entity(se, pse) == 1) {
3755 * Bias pick_next to pick the sched entity that is
3756 * triggering this preemption.
3758 if (!next_buddy_marked)
3759 set_next_buddy(pse);
3768 * Only set the backward buddy when the current task is still
3769 * on the rq. This can happen when a wakeup gets interleaved
3770 * with schedule on the ->pre_schedule() or idle_balance()
3771 * point, either of which can * drop the rq lock.
3773 * Also, during early boot the idle thread is in the fair class,
3774 * for obvious reasons its a bad idea to schedule back to it.
3776 if (unlikely(!se->on_rq || curr == rq->idle))
3779 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3783 static struct task_struct *pick_next_task_fair(struct rq *rq)
3785 struct task_struct *p;
3786 struct cfs_rq *cfs_rq = &rq->cfs;
3787 struct sched_entity *se;
3789 if (!cfs_rq->nr_running)
3793 se = pick_next_entity(cfs_rq);
3794 set_next_entity(cfs_rq, se);
3795 cfs_rq = group_cfs_rq(se);
3799 if (hrtick_enabled(rq))
3800 hrtick_start_fair(rq, p);
3806 * Account for a descheduled task:
3808 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3810 struct sched_entity *se = &prev->se;
3811 struct cfs_rq *cfs_rq;
3813 for_each_sched_entity(se) {
3814 cfs_rq = cfs_rq_of(se);
3815 put_prev_entity(cfs_rq, se);
3820 * sched_yield() is very simple
3822 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3824 static void yield_task_fair(struct rq *rq)
3826 struct task_struct *curr = rq->curr;
3827 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3828 struct sched_entity *se = &curr->se;
3831 * Are we the only task in the tree?
3833 if (unlikely(rq->nr_running == 1))
3836 clear_buddies(cfs_rq, se);
3838 if (curr->policy != SCHED_BATCH) {
3839 update_rq_clock(rq);
3841 * Update run-time statistics of the 'current'.
3843 update_curr(cfs_rq);
3845 * Tell update_rq_clock() that we've just updated,
3846 * so we don't do microscopic update in schedule()
3847 * and double the fastpath cost.
3849 rq->skip_clock_update = 1;
3855 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3857 struct sched_entity *se = &p->se;
3859 /* throttled hierarchies are not runnable */
3860 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3863 /* Tell the scheduler that we'd really like pse to run next. */
3866 yield_task_fair(rq);
3872 /**************************************************
3873 * Fair scheduling class load-balancing methods.
3877 * The purpose of load-balancing is to achieve the same basic fairness the
3878 * per-cpu scheduler provides, namely provide a proportional amount of compute
3879 * time to each task. This is expressed in the following equation:
3881 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3883 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3884 * W_i,0 is defined as:
3886 * W_i,0 = \Sum_j w_i,j (2)
3888 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3889 * is derived from the nice value as per prio_to_weight[].
3891 * The weight average is an exponential decay average of the instantaneous
3894 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3896 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3897 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3898 * can also include other factors [XXX].
3900 * To achieve this balance we define a measure of imbalance which follows
3901 * directly from (1):
3903 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3905 * We them move tasks around to minimize the imbalance. In the continuous
3906 * function space it is obvious this converges, in the discrete case we get
3907 * a few fun cases generally called infeasible weight scenarios.
3910 * - infeasible weights;
3911 * - local vs global optima in the discrete case. ]
3916 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3917 * for all i,j solution, we create a tree of cpus that follows the hardware
3918 * topology where each level pairs two lower groups (or better). This results
3919 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3920 * tree to only the first of the previous level and we decrease the frequency
3921 * of load-balance at each level inv. proportional to the number of cpus in
3927 * \Sum { --- * --- * 2^i } = O(n) (5)
3929 * `- size of each group
3930 * | | `- number of cpus doing load-balance
3932 * `- sum over all levels
3934 * Coupled with a limit on how many tasks we can migrate every balance pass,
3935 * this makes (5) the runtime complexity of the balancer.
3937 * An important property here is that each CPU is still (indirectly) connected
3938 * to every other cpu in at most O(log n) steps:
3940 * The adjacency matrix of the resulting graph is given by:
3943 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3946 * And you'll find that:
3948 * A^(log_2 n)_i,j != 0 for all i,j (7)
3950 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3951 * The task movement gives a factor of O(m), giving a convergence complexity
3954 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3959 * In order to avoid CPUs going idle while there's still work to do, new idle
3960 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3961 * tree itself instead of relying on other CPUs to bring it work.
3963 * This adds some complexity to both (5) and (8) but it reduces the total idle
3971 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3974 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3979 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3981 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3983 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3986 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3987 * rewrite all of this once again.]
3990 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3992 #define LBF_ALL_PINNED 0x01
3993 #define LBF_NEED_BREAK 0x02
3994 #define LBF_DST_PINNED 0x04
3995 #define LBF_SOME_PINNED 0x08
3998 struct sched_domain *sd;
4006 struct cpumask *dst_grpmask;
4008 enum cpu_idle_type idle;
4010 /* The set of CPUs under consideration for load-balancing */
4011 struct cpumask *cpus;
4016 unsigned int loop_break;
4017 unsigned int loop_max;
4021 * move_task - move a task from one runqueue to another runqueue.
4022 * Both runqueues must be locked.
4024 static void move_task(struct task_struct *p, struct lb_env *env)
4026 deactivate_task(env->src_rq, p, 0);
4027 set_task_cpu(p, env->dst_cpu);
4028 activate_task(env->dst_rq, p, 0);
4029 check_preempt_curr(env->dst_rq, p, 0);
4033 * Is this task likely cache-hot:
4036 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4040 if (p->sched_class != &fair_sched_class)
4043 if (unlikely(p->policy == SCHED_IDLE))
4047 * Buddy candidates are cache hot:
4049 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4050 (&p->se == cfs_rq_of(&p->se)->next ||
4051 &p->se == cfs_rq_of(&p->se)->last))
4054 if (sysctl_sched_migration_cost == -1)
4056 if (sysctl_sched_migration_cost == 0)
4059 delta = now - p->se.exec_start;
4061 return delta < (s64)sysctl_sched_migration_cost;
4065 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4068 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4070 int tsk_cache_hot = 0;
4072 * We do not migrate tasks that are:
4073 * 1) throttled_lb_pair, or
4074 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4075 * 3) running (obviously), or
4076 * 4) are cache-hot on their current CPU.
4078 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4081 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4084 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4086 env->flags |= LBF_SOME_PINNED;
4089 * Remember if this task can be migrated to any other cpu in
4090 * our sched_group. We may want to revisit it if we couldn't
4091 * meet load balance goals by pulling other tasks on src_cpu.
4093 * Also avoid computing new_dst_cpu if we have already computed
4094 * one in current iteration.
4096 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4099 /* Prevent to re-select dst_cpu via env's cpus */
4100 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4101 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4102 env->flags |= LBF_DST_PINNED;
4103 env->new_dst_cpu = cpu;
4111 /* Record that we found atleast one task that could run on dst_cpu */
4112 env->flags &= ~LBF_ALL_PINNED;
4114 if (task_running(env->src_rq, p)) {
4115 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4120 * Aggressive migration if:
4121 * 1) task is cache cold, or
4122 * 2) too many balance attempts have failed.
4125 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4126 if (!tsk_cache_hot ||
4127 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4129 if (tsk_cache_hot) {
4130 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4131 schedstat_inc(p, se.statistics.nr_forced_migrations);
4137 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4142 * move_one_task tries to move exactly one task from busiest to this_rq, as
4143 * part of active balancing operations within "domain".
4144 * Returns 1 if successful and 0 otherwise.
4146 * Called with both runqueues locked.
4148 static int move_one_task(struct lb_env *env)
4150 struct task_struct *p, *n;
4152 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4153 if (!can_migrate_task(p, env))
4158 * Right now, this is only the second place move_task()
4159 * is called, so we can safely collect move_task()
4160 * stats here rather than inside move_task().
4162 schedstat_inc(env->sd, lb_gained[env->idle]);
4168 static unsigned long task_h_load(struct task_struct *p);
4170 static const unsigned int sched_nr_migrate_break = 32;
4173 * move_tasks tries to move up to imbalance weighted load from busiest to
4174 * this_rq, as part of a balancing operation within domain "sd".
4175 * Returns 1 if successful and 0 otherwise.
4177 * Called with both runqueues locked.
4179 static int move_tasks(struct lb_env *env)
4181 struct list_head *tasks = &env->src_rq->cfs_tasks;
4182 struct task_struct *p;
4186 if (env->imbalance <= 0)
4189 while (!list_empty(tasks)) {
4190 p = list_first_entry(tasks, struct task_struct, se.group_node);
4193 /* We've more or less seen every task there is, call it quits */
4194 if (env->loop > env->loop_max)
4197 /* take a breather every nr_migrate tasks */
4198 if (env->loop > env->loop_break) {
4199 env->loop_break += sched_nr_migrate_break;
4200 env->flags |= LBF_NEED_BREAK;
4204 if (!can_migrate_task(p, env))
4207 load = task_h_load(p);
4209 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4212 if ((load / 2) > env->imbalance)
4217 env->imbalance -= load;
4219 #ifdef CONFIG_PREEMPT
4221 * NEWIDLE balancing is a source of latency, so preemptible
4222 * kernels will stop after the first task is pulled to minimize
4223 * the critical section.
4225 if (env->idle == CPU_NEWLY_IDLE)
4230 * We only want to steal up to the prescribed amount of
4233 if (env->imbalance <= 0)
4238 list_move_tail(&p->se.group_node, tasks);
4242 * Right now, this is one of only two places move_task() is called,
4243 * so we can safely collect move_task() stats here rather than
4244 * inside move_task().
4246 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4251 #ifdef CONFIG_FAIR_GROUP_SCHED
4253 * update tg->load_weight by folding this cpu's load_avg
4255 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4257 struct sched_entity *se = tg->se[cpu];
4258 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4260 /* throttled entities do not contribute to load */
4261 if (throttled_hierarchy(cfs_rq))
4264 update_cfs_rq_blocked_load(cfs_rq, 1);
4267 update_entity_load_avg(se, 1);
4269 * We pivot on our runnable average having decayed to zero for
4270 * list removal. This generally implies that all our children
4271 * have also been removed (modulo rounding error or bandwidth
4272 * control); however, such cases are rare and we can fix these
4275 * TODO: fix up out-of-order children on enqueue.
4277 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4278 list_del_leaf_cfs_rq(cfs_rq);
4280 struct rq *rq = rq_of(cfs_rq);
4281 update_rq_runnable_avg(rq, rq->nr_running);
4285 static void update_blocked_averages(int cpu)
4287 struct rq *rq = cpu_rq(cpu);
4288 struct cfs_rq *cfs_rq;
4289 unsigned long flags;
4291 raw_spin_lock_irqsave(&rq->lock, flags);
4292 update_rq_clock(rq);
4294 * Iterates the task_group tree in a bottom up fashion, see
4295 * list_add_leaf_cfs_rq() for details.
4297 for_each_leaf_cfs_rq(rq, cfs_rq) {
4299 * Note: We may want to consider periodically releasing
4300 * rq->lock about these updates so that creating many task
4301 * groups does not result in continually extending hold time.
4303 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4306 raw_spin_unlock_irqrestore(&rq->lock, flags);
4310 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4311 * This needs to be done in a top-down fashion because the load of a child
4312 * group is a fraction of its parents load.
4314 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4316 struct rq *rq = rq_of(cfs_rq);
4317 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4318 unsigned long now = jiffies;
4321 if (cfs_rq->last_h_load_update == now)
4324 cfs_rq->h_load_next = NULL;
4325 for_each_sched_entity(se) {
4326 cfs_rq = cfs_rq_of(se);
4327 cfs_rq->h_load_next = se;
4328 if (cfs_rq->last_h_load_update == now)
4333 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4334 cfs_rq->last_h_load_update = now;
4337 while ((se = cfs_rq->h_load_next) != NULL) {
4338 load = cfs_rq->h_load;
4339 load = div64_ul(load * se->avg.load_avg_contrib,
4340 cfs_rq->runnable_load_avg + 1);
4341 cfs_rq = group_cfs_rq(se);
4342 cfs_rq->h_load = load;
4343 cfs_rq->last_h_load_update = now;
4347 static unsigned long task_h_load(struct task_struct *p)
4349 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4351 update_cfs_rq_h_load(cfs_rq);
4352 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4353 cfs_rq->runnable_load_avg + 1);
4356 static inline void update_blocked_averages(int cpu)
4360 static unsigned long task_h_load(struct task_struct *p)
4362 return p->se.avg.load_avg_contrib;
4366 /********** Helpers for find_busiest_group ************************/
4368 * sg_lb_stats - stats of a sched_group required for load_balancing
4370 struct sg_lb_stats {
4371 unsigned long avg_load; /*Avg load across the CPUs of the group */
4372 unsigned long group_load; /* Total load over the CPUs of the group */
4373 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4374 unsigned long load_per_task;
4375 unsigned long group_power;
4376 unsigned int sum_nr_running; /* Nr tasks running in the group */
4377 unsigned int group_capacity;
4378 unsigned int idle_cpus;
4379 unsigned int group_weight;
4380 int group_imb; /* Is there an imbalance in the group ? */
4381 int group_has_capacity; /* Is there extra capacity in the group? */
4385 * sd_lb_stats - Structure to store the statistics of a sched_domain
4386 * during load balancing.
4388 struct sd_lb_stats {
4389 struct sched_group *busiest; /* Busiest group in this sd */
4390 struct sched_group *local; /* Local group in this sd */
4391 unsigned long total_load; /* Total load of all groups in sd */
4392 unsigned long total_pwr; /* Total power of all groups in sd */
4393 unsigned long avg_load; /* Average load across all groups in sd */
4395 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4396 struct sg_lb_stats local_stat; /* Statistics of the local group */
4399 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4402 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4403 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4404 * We must however clear busiest_stat::avg_load because
4405 * update_sd_pick_busiest() reads this before assignment.
4407 *sds = (struct sd_lb_stats){
4419 * get_sd_load_idx - Obtain the load index for a given sched domain.
4420 * @sd: The sched_domain whose load_idx is to be obtained.
4421 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4423 * Return: The load index.
4425 static inline int get_sd_load_idx(struct sched_domain *sd,
4426 enum cpu_idle_type idle)
4432 load_idx = sd->busy_idx;
4435 case CPU_NEWLY_IDLE:
4436 load_idx = sd->newidle_idx;
4439 load_idx = sd->idle_idx;
4446 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4448 return SCHED_POWER_SCALE;
4451 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4453 return default_scale_freq_power(sd, cpu);
4456 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4458 unsigned long weight = sd->span_weight;
4459 unsigned long smt_gain = sd->smt_gain;
4466 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4468 return default_scale_smt_power(sd, cpu);
4471 static unsigned long scale_rt_power(int cpu)
4473 struct rq *rq = cpu_rq(cpu);
4474 u64 total, available, age_stamp, avg;
4477 * Since we're reading these variables without serialization make sure
4478 * we read them once before doing sanity checks on them.
4480 age_stamp = ACCESS_ONCE(rq->age_stamp);
4481 avg = ACCESS_ONCE(rq->rt_avg);
4483 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4485 if (unlikely(total < avg)) {
4486 /* Ensures that power won't end up being negative */
4489 available = total - avg;
4492 if (unlikely((s64)total < SCHED_POWER_SCALE))
4493 total = SCHED_POWER_SCALE;
4495 total >>= SCHED_POWER_SHIFT;
4497 return div_u64(available, total);
4500 static void update_cpu_power(struct sched_domain *sd, int cpu)
4502 unsigned long weight = sd->span_weight;
4503 unsigned long power = SCHED_POWER_SCALE;
4504 struct sched_group *sdg = sd->groups;
4506 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4507 if (sched_feat(ARCH_POWER))
4508 power *= arch_scale_smt_power(sd, cpu);
4510 power *= default_scale_smt_power(sd, cpu);
4512 power >>= SCHED_POWER_SHIFT;
4515 sdg->sgp->power_orig = power;
4517 if (sched_feat(ARCH_POWER))
4518 power *= arch_scale_freq_power(sd, cpu);
4520 power *= default_scale_freq_power(sd, cpu);
4522 power >>= SCHED_POWER_SHIFT;
4524 power *= scale_rt_power(cpu);
4525 power >>= SCHED_POWER_SHIFT;
4530 cpu_rq(cpu)->cpu_power = power;
4531 sdg->sgp->power = power;
4534 void update_group_power(struct sched_domain *sd, int cpu)
4536 struct sched_domain *child = sd->child;
4537 struct sched_group *group, *sdg = sd->groups;
4538 unsigned long power, power_orig;
4539 unsigned long interval;
4541 interval = msecs_to_jiffies(sd->balance_interval);
4542 interval = clamp(interval, 1UL, max_load_balance_interval);
4543 sdg->sgp->next_update = jiffies + interval;
4546 update_cpu_power(sd, cpu);
4550 power_orig = power = 0;
4552 if (child->flags & SD_OVERLAP) {
4554 * SD_OVERLAP domains cannot assume that child groups
4555 * span the current group.
4558 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4559 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4561 power_orig += sg->sgp->power_orig;
4562 power += sg->sgp->power;
4566 * !SD_OVERLAP domains can assume that child groups
4567 * span the current group.
4570 group = child->groups;
4572 power_orig += group->sgp->power_orig;
4573 power += group->sgp->power;
4574 group = group->next;
4575 } while (group != child->groups);
4578 sdg->sgp->power_orig = power_orig;
4579 sdg->sgp->power = power;
4583 * Try and fix up capacity for tiny siblings, this is needed when
4584 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4585 * which on its own isn't powerful enough.
4587 * See update_sd_pick_busiest() and check_asym_packing().
4590 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4593 * Only siblings can have significantly less than SCHED_POWER_SCALE
4595 if (!(sd->flags & SD_SHARE_CPUPOWER))
4599 * If ~90% of the cpu_power is still there, we're good.
4601 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4608 * Group imbalance indicates (and tries to solve) the problem where balancing
4609 * groups is inadequate due to tsk_cpus_allowed() constraints.
4611 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4612 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4615 * { 0 1 2 3 } { 4 5 6 7 }
4618 * If we were to balance group-wise we'd place two tasks in the first group and
4619 * two tasks in the second group. Clearly this is undesired as it will overload
4620 * cpu 3 and leave one of the cpus in the second group unused.
4622 * The current solution to this issue is detecting the skew in the first group
4623 * by noticing the lower domain failed to reach balance and had difficulty
4624 * moving tasks due to affinity constraints.
4626 * When this is so detected; this group becomes a candidate for busiest; see
4627 * update_sd_pick_busiest(). And calculcate_imbalance() and
4628 * find_busiest_group() avoid some of the usual balance conditions to allow it
4629 * to create an effective group imbalance.
4631 * This is a somewhat tricky proposition since the next run might not find the
4632 * group imbalance and decide the groups need to be balanced again. A most
4633 * subtle and fragile situation.
4636 static inline int sg_imbalanced(struct sched_group *group)
4638 return group->sgp->imbalance;
4642 * Compute the group capacity.
4644 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4645 * first dividing out the smt factor and computing the actual number of cores
4646 * and limit power unit capacity with that.
4648 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4650 unsigned int capacity, smt, cpus;
4651 unsigned int power, power_orig;
4653 power = group->sgp->power;
4654 power_orig = group->sgp->power_orig;
4655 cpus = group->group_weight;
4657 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4658 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4659 capacity = cpus / smt; /* cores */
4661 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4663 capacity = fix_small_capacity(env->sd, group);
4669 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4670 * @env: The load balancing environment.
4671 * @group: sched_group whose statistics are to be updated.
4672 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4673 * @local_group: Does group contain this_cpu.
4674 * @sgs: variable to hold the statistics for this group.
4676 static inline void update_sg_lb_stats(struct lb_env *env,
4677 struct sched_group *group, int load_idx,
4678 int local_group, struct sg_lb_stats *sgs)
4680 unsigned long nr_running;
4684 memset(sgs, 0, sizeof(*sgs));
4686 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4687 struct rq *rq = cpu_rq(i);
4689 nr_running = rq->nr_running;
4691 /* Bias balancing toward cpus of our domain */
4693 load = target_load(i, load_idx);
4695 load = source_load(i, load_idx);
4697 sgs->group_load += load;
4698 sgs->sum_nr_running += nr_running;
4699 sgs->sum_weighted_load += weighted_cpuload(i);
4704 /* Adjust by relative CPU power of the group */
4705 sgs->group_power = group->sgp->power;
4706 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4708 if (sgs->sum_nr_running)
4709 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4711 sgs->group_weight = group->group_weight;
4713 sgs->group_imb = sg_imbalanced(group);
4714 sgs->group_capacity = sg_capacity(env, group);
4716 if (sgs->group_capacity > sgs->sum_nr_running)
4717 sgs->group_has_capacity = 1;
4721 * update_sd_pick_busiest - return 1 on busiest group
4722 * @env: The load balancing environment.
4723 * @sds: sched_domain statistics
4724 * @sg: sched_group candidate to be checked for being the busiest
4725 * @sgs: sched_group statistics
4727 * Determine if @sg is a busier group than the previously selected
4730 * Return: %true if @sg is a busier group than the previously selected
4731 * busiest group. %false otherwise.
4733 static bool update_sd_pick_busiest(struct lb_env *env,
4734 struct sd_lb_stats *sds,
4735 struct sched_group *sg,
4736 struct sg_lb_stats *sgs)
4738 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4741 if (sgs->sum_nr_running > sgs->group_capacity)
4748 * ASYM_PACKING needs to move all the work to the lowest
4749 * numbered CPUs in the group, therefore mark all groups
4750 * higher than ourself as busy.
4752 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4753 env->dst_cpu < group_first_cpu(sg)) {
4757 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4765 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4766 * @env: The load balancing environment.
4767 * @balance: Should we balance.
4768 * @sds: variable to hold the statistics for this sched_domain.
4770 static inline void update_sd_lb_stats(struct lb_env *env,
4771 struct sd_lb_stats *sds)
4773 struct sched_domain *child = env->sd->child;
4774 struct sched_group *sg = env->sd->groups;
4775 struct sg_lb_stats tmp_sgs;
4776 int load_idx, prefer_sibling = 0;
4778 if (child && child->flags & SD_PREFER_SIBLING)
4781 load_idx = get_sd_load_idx(env->sd, env->idle);
4784 struct sg_lb_stats *sgs = &tmp_sgs;
4787 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4790 sgs = &sds->local_stat;
4792 if (env->idle != CPU_NEWLY_IDLE ||
4793 time_after_eq(jiffies, sg->sgp->next_update))
4794 update_group_power(env->sd, env->dst_cpu);
4797 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4803 * In case the child domain prefers tasks go to siblings
4804 * first, lower the sg capacity to one so that we'll try
4805 * and move all the excess tasks away. We lower the capacity
4806 * of a group only if the local group has the capacity to fit
4807 * these excess tasks, i.e. nr_running < group_capacity. The
4808 * extra check prevents the case where you always pull from the
4809 * heaviest group when it is already under-utilized (possible
4810 * with a large weight task outweighs the tasks on the system).
4812 if (prefer_sibling && sds->local &&
4813 sds->local_stat.group_has_capacity)
4814 sgs->group_capacity = min(sgs->group_capacity, 1U);
4816 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4818 sds->busiest_stat = *sgs;
4822 /* Now, start updating sd_lb_stats */
4823 sds->total_load += sgs->group_load;
4824 sds->total_pwr += sgs->group_power;
4827 } while (sg != env->sd->groups);
4831 * check_asym_packing - Check to see if the group is packed into the
4834 * This is primarily intended to used at the sibling level. Some
4835 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4836 * case of POWER7, it can move to lower SMT modes only when higher
4837 * threads are idle. When in lower SMT modes, the threads will
4838 * perform better since they share less core resources. Hence when we
4839 * have idle threads, we want them to be the higher ones.
4841 * This packing function is run on idle threads. It checks to see if
4842 * the busiest CPU in this domain (core in the P7 case) has a higher
4843 * CPU number than the packing function is being run on. Here we are
4844 * assuming lower CPU number will be equivalent to lower a SMT thread
4847 * Return: 1 when packing is required and a task should be moved to
4848 * this CPU. The amount of the imbalance is returned in *imbalance.
4850 * @env: The load balancing environment.
4851 * @sds: Statistics of the sched_domain which is to be packed
4853 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4857 if (!(env->sd->flags & SD_ASYM_PACKING))
4863 busiest_cpu = group_first_cpu(sds->busiest);
4864 if (env->dst_cpu > busiest_cpu)
4867 env->imbalance = DIV_ROUND_CLOSEST(
4868 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4875 * fix_small_imbalance - Calculate the minor imbalance that exists
4876 * amongst the groups of a sched_domain, during
4878 * @env: The load balancing environment.
4879 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4882 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4884 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4885 unsigned int imbn = 2;
4886 unsigned long scaled_busy_load_per_task;
4887 struct sg_lb_stats *local, *busiest;
4889 local = &sds->local_stat;
4890 busiest = &sds->busiest_stat;
4892 if (!local->sum_nr_running)
4893 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4894 else if (busiest->load_per_task > local->load_per_task)
4897 scaled_busy_load_per_task =
4898 (busiest->load_per_task * SCHED_POWER_SCALE) /
4899 busiest->group_power;
4901 if (busiest->avg_load + scaled_busy_load_per_task >=
4902 local->avg_load + (scaled_busy_load_per_task * imbn)) {
4903 env->imbalance = busiest->load_per_task;
4908 * OK, we don't have enough imbalance to justify moving tasks,
4909 * however we may be able to increase total CPU power used by
4913 pwr_now += busiest->group_power *
4914 min(busiest->load_per_task, busiest->avg_load);
4915 pwr_now += local->group_power *
4916 min(local->load_per_task, local->avg_load);
4917 pwr_now /= SCHED_POWER_SCALE;
4919 /* Amount of load we'd subtract */
4920 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4921 busiest->group_power;
4922 if (busiest->avg_load > tmp) {
4923 pwr_move += busiest->group_power *
4924 min(busiest->load_per_task,
4925 busiest->avg_load - tmp);
4928 /* Amount of load we'd add */
4929 if (busiest->avg_load * busiest->group_power <
4930 busiest->load_per_task * SCHED_POWER_SCALE) {
4931 tmp = (busiest->avg_load * busiest->group_power) /
4934 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
4937 pwr_move += local->group_power *
4938 min(local->load_per_task, local->avg_load + tmp);
4939 pwr_move /= SCHED_POWER_SCALE;
4941 /* Move if we gain throughput */
4942 if (pwr_move > pwr_now)
4943 env->imbalance = busiest->load_per_task;
4947 * calculate_imbalance - Calculate the amount of imbalance present within the
4948 * groups of a given sched_domain during load balance.
4949 * @env: load balance environment
4950 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4952 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4954 unsigned long max_pull, load_above_capacity = ~0UL;
4955 struct sg_lb_stats *local, *busiest;
4957 local = &sds->local_stat;
4958 busiest = &sds->busiest_stat;
4960 if (busiest->group_imb) {
4962 * In the group_imb case we cannot rely on group-wide averages
4963 * to ensure cpu-load equilibrium, look at wider averages. XXX
4965 busiest->load_per_task =
4966 min(busiest->load_per_task, sds->avg_load);
4970 * In the presence of smp nice balancing, certain scenarios can have
4971 * max load less than avg load(as we skip the groups at or below
4972 * its cpu_power, while calculating max_load..)
4974 if (busiest->avg_load <= sds->avg_load ||
4975 local->avg_load >= sds->avg_load) {
4977 return fix_small_imbalance(env, sds);
4980 if (!busiest->group_imb) {
4982 * Don't want to pull so many tasks that a group would go idle.
4983 * Except of course for the group_imb case, since then we might
4984 * have to drop below capacity to reach cpu-load equilibrium.
4986 load_above_capacity =
4987 (busiest->sum_nr_running - busiest->group_capacity);
4989 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4990 load_above_capacity /= busiest->group_power;
4994 * We're trying to get all the cpus to the average_load, so we don't
4995 * want to push ourselves above the average load, nor do we wish to
4996 * reduce the max loaded cpu below the average load. At the same time,
4997 * we also don't want to reduce the group load below the group capacity
4998 * (so that we can implement power-savings policies etc). Thus we look
4999 * for the minimum possible imbalance.
5001 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5003 /* How much load to actually move to equalise the imbalance */
5004 env->imbalance = min(
5005 max_pull * busiest->group_power,
5006 (sds->avg_load - local->avg_load) * local->group_power
5007 ) / SCHED_POWER_SCALE;
5010 * if *imbalance is less than the average load per runnable task
5011 * there is no guarantee that any tasks will be moved so we'll have
5012 * a think about bumping its value to force at least one task to be
5015 if (env->imbalance < busiest->load_per_task)
5016 return fix_small_imbalance(env, sds);
5019 /******* find_busiest_group() helpers end here *********************/
5022 * find_busiest_group - Returns the busiest group within the sched_domain
5023 * if there is an imbalance. If there isn't an imbalance, and
5024 * the user has opted for power-savings, it returns a group whose
5025 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5026 * such a group exists.
5028 * Also calculates the amount of weighted load which should be moved
5029 * to restore balance.
5031 * @env: The load balancing environment.
5033 * Return: - The busiest group if imbalance exists.
5034 * - If no imbalance and user has opted for power-savings balance,
5035 * return the least loaded group whose CPUs can be
5036 * put to idle by rebalancing its tasks onto our group.
5038 static struct sched_group *find_busiest_group(struct lb_env *env)
5040 struct sg_lb_stats *local, *busiest;
5041 struct sd_lb_stats sds;
5043 init_sd_lb_stats(&sds);
5046 * Compute the various statistics relavent for load balancing at
5049 update_sd_lb_stats(env, &sds);
5050 local = &sds.local_stat;
5051 busiest = &sds.busiest_stat;
5053 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5054 check_asym_packing(env, &sds))
5057 /* There is no busy sibling group to pull tasks from */
5058 if (!sds.busiest || busiest->sum_nr_running == 0)
5061 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5064 * If the busiest group is imbalanced the below checks don't
5065 * work because they assume all things are equal, which typically
5066 * isn't true due to cpus_allowed constraints and the like.
5068 if (busiest->group_imb)
5071 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5072 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5073 !busiest->group_has_capacity)
5077 * If the local group is more busy than the selected busiest group
5078 * don't try and pull any tasks.
5080 if (local->avg_load >= busiest->avg_load)
5084 * Don't pull any tasks if this group is already above the domain
5087 if (local->avg_load >= sds.avg_load)
5090 if (env->idle == CPU_IDLE) {
5092 * This cpu is idle. If the busiest group load doesn't
5093 * have more tasks than the number of available cpu's and
5094 * there is no imbalance between this and busiest group
5095 * wrt to idle cpu's, it is balanced.
5097 if ((local->idle_cpus < busiest->idle_cpus) &&
5098 busiest->sum_nr_running <= busiest->group_weight)
5102 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5103 * imbalance_pct to be conservative.
5105 if (100 * busiest->avg_load <=
5106 env->sd->imbalance_pct * local->avg_load)
5111 /* Looks like there is an imbalance. Compute it */
5112 calculate_imbalance(env, &sds);
5121 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5123 static struct rq *find_busiest_queue(struct lb_env *env,
5124 struct sched_group *group)
5126 struct rq *busiest = NULL, *rq;
5127 unsigned long busiest_load = 0, busiest_power = 1;
5130 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5131 unsigned long power = power_of(i);
5132 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5137 capacity = fix_small_capacity(env->sd, group);
5140 wl = weighted_cpuload(i);
5143 * When comparing with imbalance, use weighted_cpuload()
5144 * which is not scaled with the cpu power.
5146 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5150 * For the load comparisons with the other cpu's, consider
5151 * the weighted_cpuload() scaled with the cpu power, so that
5152 * the load can be moved away from the cpu that is potentially
5153 * running at a lower capacity.
5155 * Thus we're looking for max(wl_i / power_i), crosswise
5156 * multiplication to rid ourselves of the division works out
5157 * to: wl_i * power_j > wl_j * power_i; where j is our
5160 if (wl * busiest_power > busiest_load * power) {
5162 busiest_power = power;
5171 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5172 * so long as it is large enough.
5174 #define MAX_PINNED_INTERVAL 512
5176 /* Working cpumask for load_balance and load_balance_newidle. */
5177 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5179 static int need_active_balance(struct lb_env *env)
5181 struct sched_domain *sd = env->sd;
5183 if (env->idle == CPU_NEWLY_IDLE) {
5186 * ASYM_PACKING needs to force migrate tasks from busy but
5187 * higher numbered CPUs in order to pack all tasks in the
5188 * lowest numbered CPUs.
5190 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5194 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5197 static int active_load_balance_cpu_stop(void *data);
5199 static int should_we_balance(struct lb_env *env)
5201 struct sched_group *sg = env->sd->groups;
5202 struct cpumask *sg_cpus, *sg_mask;
5203 int cpu, balance_cpu = -1;
5206 * In the newly idle case, we will allow all the cpu's
5207 * to do the newly idle load balance.
5209 if (env->idle == CPU_NEWLY_IDLE)
5212 sg_cpus = sched_group_cpus(sg);
5213 sg_mask = sched_group_mask(sg);
5214 /* Try to find first idle cpu */
5215 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5216 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5223 if (balance_cpu == -1)
5224 balance_cpu = group_balance_cpu(sg);
5227 * First idle cpu or the first cpu(busiest) in this sched group
5228 * is eligible for doing load balancing at this and above domains.
5230 return balance_cpu == env->dst_cpu;
5234 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5235 * tasks if there is an imbalance.
5237 static int load_balance(int this_cpu, struct rq *this_rq,
5238 struct sched_domain *sd, enum cpu_idle_type idle,
5239 int *continue_balancing)
5241 int ld_moved, cur_ld_moved, active_balance = 0;
5242 struct sched_domain *sd_parent = sd->parent;
5243 struct sched_group *group;
5245 unsigned long flags;
5246 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5248 struct lb_env env = {
5250 .dst_cpu = this_cpu,
5252 .dst_grpmask = sched_group_cpus(sd->groups),
5254 .loop_break = sched_nr_migrate_break,
5259 * For NEWLY_IDLE load_balancing, we don't need to consider
5260 * other cpus in our group
5262 if (idle == CPU_NEWLY_IDLE)
5263 env.dst_grpmask = NULL;
5265 cpumask_copy(cpus, cpu_active_mask);
5267 schedstat_inc(sd, lb_count[idle]);
5270 if (!should_we_balance(&env)) {
5271 *continue_balancing = 0;
5275 group = find_busiest_group(&env);
5277 schedstat_inc(sd, lb_nobusyg[idle]);
5281 busiest = find_busiest_queue(&env, group);
5283 schedstat_inc(sd, lb_nobusyq[idle]);
5287 BUG_ON(busiest == env.dst_rq);
5289 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5292 if (busiest->nr_running > 1) {
5294 * Attempt to move tasks. If find_busiest_group has found
5295 * an imbalance but busiest->nr_running <= 1, the group is
5296 * still unbalanced. ld_moved simply stays zero, so it is
5297 * correctly treated as an imbalance.
5299 env.flags |= LBF_ALL_PINNED;
5300 env.src_cpu = busiest->cpu;
5301 env.src_rq = busiest;
5302 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5305 local_irq_save(flags);
5306 double_rq_lock(env.dst_rq, busiest);
5309 * cur_ld_moved - load moved in current iteration
5310 * ld_moved - cumulative load moved across iterations
5312 cur_ld_moved = move_tasks(&env);
5313 ld_moved += cur_ld_moved;
5314 double_rq_unlock(env.dst_rq, busiest);
5315 local_irq_restore(flags);
5318 * some other cpu did the load balance for us.
5320 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5321 resched_cpu(env.dst_cpu);
5323 if (env.flags & LBF_NEED_BREAK) {
5324 env.flags &= ~LBF_NEED_BREAK;
5329 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5330 * us and move them to an alternate dst_cpu in our sched_group
5331 * where they can run. The upper limit on how many times we
5332 * iterate on same src_cpu is dependent on number of cpus in our
5335 * This changes load balance semantics a bit on who can move
5336 * load to a given_cpu. In addition to the given_cpu itself
5337 * (or a ilb_cpu acting on its behalf where given_cpu is
5338 * nohz-idle), we now have balance_cpu in a position to move
5339 * load to given_cpu. In rare situations, this may cause
5340 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5341 * _independently_ and at _same_ time to move some load to
5342 * given_cpu) causing exceess load to be moved to given_cpu.
5343 * This however should not happen so much in practice and
5344 * moreover subsequent load balance cycles should correct the
5345 * excess load moved.
5347 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5349 /* Prevent to re-select dst_cpu via env's cpus */
5350 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5352 env.dst_rq = cpu_rq(env.new_dst_cpu);
5353 env.dst_cpu = env.new_dst_cpu;
5354 env.flags &= ~LBF_DST_PINNED;
5356 env.loop_break = sched_nr_migrate_break;
5359 * Go back to "more_balance" rather than "redo" since we
5360 * need to continue with same src_cpu.
5366 * We failed to reach balance because of affinity.
5369 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5371 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5372 *group_imbalance = 1;
5373 } else if (*group_imbalance)
5374 *group_imbalance = 0;
5377 /* All tasks on this runqueue were pinned by CPU affinity */
5378 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5379 cpumask_clear_cpu(cpu_of(busiest), cpus);
5380 if (!cpumask_empty(cpus)) {
5382 env.loop_break = sched_nr_migrate_break;
5390 schedstat_inc(sd, lb_failed[idle]);
5392 * Increment the failure counter only on periodic balance.
5393 * We do not want newidle balance, which can be very
5394 * frequent, pollute the failure counter causing
5395 * excessive cache_hot migrations and active balances.
5397 if (idle != CPU_NEWLY_IDLE)
5398 sd->nr_balance_failed++;
5400 if (need_active_balance(&env)) {
5401 raw_spin_lock_irqsave(&busiest->lock, flags);
5403 /* don't kick the active_load_balance_cpu_stop,
5404 * if the curr task on busiest cpu can't be
5407 if (!cpumask_test_cpu(this_cpu,
5408 tsk_cpus_allowed(busiest->curr))) {
5409 raw_spin_unlock_irqrestore(&busiest->lock,
5411 env.flags |= LBF_ALL_PINNED;
5412 goto out_one_pinned;
5416 * ->active_balance synchronizes accesses to
5417 * ->active_balance_work. Once set, it's cleared
5418 * only after active load balance is finished.
5420 if (!busiest->active_balance) {
5421 busiest->active_balance = 1;
5422 busiest->push_cpu = this_cpu;
5425 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5427 if (active_balance) {
5428 stop_one_cpu_nowait(cpu_of(busiest),
5429 active_load_balance_cpu_stop, busiest,
5430 &busiest->active_balance_work);
5434 * We've kicked active balancing, reset the failure
5437 sd->nr_balance_failed = sd->cache_nice_tries+1;
5440 sd->nr_balance_failed = 0;
5442 if (likely(!active_balance)) {
5443 /* We were unbalanced, so reset the balancing interval */
5444 sd->balance_interval = sd->min_interval;
5447 * If we've begun active balancing, start to back off. This
5448 * case may not be covered by the all_pinned logic if there
5449 * is only 1 task on the busy runqueue (because we don't call
5452 if (sd->balance_interval < sd->max_interval)
5453 sd->balance_interval *= 2;
5459 schedstat_inc(sd, lb_balanced[idle]);
5461 sd->nr_balance_failed = 0;
5464 /* tune up the balancing interval */
5465 if (((env.flags & LBF_ALL_PINNED) &&
5466 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5467 (sd->balance_interval < sd->max_interval))
5468 sd->balance_interval *= 2;
5476 * idle_balance is called by schedule() if this_cpu is about to become
5477 * idle. Attempts to pull tasks from other CPUs.
5479 void idle_balance(int this_cpu, struct rq *this_rq)
5481 struct sched_domain *sd;
5482 int pulled_task = 0;
5483 unsigned long next_balance = jiffies + HZ;
5486 this_rq->idle_stamp = rq_clock(this_rq);
5488 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5492 * Drop the rq->lock, but keep IRQ/preempt disabled.
5494 raw_spin_unlock(&this_rq->lock);
5496 update_blocked_averages(this_cpu);
5498 for_each_domain(this_cpu, sd) {
5499 unsigned long interval;
5500 int continue_balancing = 1;
5501 u64 t0, domain_cost;
5503 if (!(sd->flags & SD_LOAD_BALANCE))
5506 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5509 if (sd->flags & SD_BALANCE_NEWIDLE) {
5510 t0 = sched_clock_cpu(this_cpu);
5512 /* If we've pulled tasks over stop searching: */
5513 pulled_task = load_balance(this_cpu, this_rq,
5515 &continue_balancing);
5517 domain_cost = sched_clock_cpu(this_cpu) - t0;
5518 if (domain_cost > sd->max_newidle_lb_cost)
5519 sd->max_newidle_lb_cost = domain_cost;
5521 curr_cost += domain_cost;
5524 interval = msecs_to_jiffies(sd->balance_interval);
5525 if (time_after(next_balance, sd->last_balance + interval))
5526 next_balance = sd->last_balance + interval;
5528 this_rq->idle_stamp = 0;
5534 raw_spin_lock(&this_rq->lock);
5536 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5538 * We are going idle. next_balance may be set based on
5539 * a busy processor. So reset next_balance.
5541 this_rq->next_balance = next_balance;
5544 if (curr_cost > this_rq->max_idle_balance_cost)
5545 this_rq->max_idle_balance_cost = curr_cost;
5549 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5550 * running tasks off the busiest CPU onto idle CPUs. It requires at
5551 * least 1 task to be running on each physical CPU where possible, and
5552 * avoids physical / logical imbalances.
5554 static int active_load_balance_cpu_stop(void *data)
5556 struct rq *busiest_rq = data;
5557 int busiest_cpu = cpu_of(busiest_rq);
5558 int target_cpu = busiest_rq->push_cpu;
5559 struct rq *target_rq = cpu_rq(target_cpu);
5560 struct sched_domain *sd;
5562 raw_spin_lock_irq(&busiest_rq->lock);
5564 /* make sure the requested cpu hasn't gone down in the meantime */
5565 if (unlikely(busiest_cpu != smp_processor_id() ||
5566 !busiest_rq->active_balance))
5569 /* Is there any task to move? */
5570 if (busiest_rq->nr_running <= 1)
5574 * This condition is "impossible", if it occurs
5575 * we need to fix it. Originally reported by
5576 * Bjorn Helgaas on a 128-cpu setup.
5578 BUG_ON(busiest_rq == target_rq);
5580 /* move a task from busiest_rq to target_rq */
5581 double_lock_balance(busiest_rq, target_rq);
5583 /* Search for an sd spanning us and the target CPU. */
5585 for_each_domain(target_cpu, sd) {
5586 if ((sd->flags & SD_LOAD_BALANCE) &&
5587 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5592 struct lb_env env = {
5594 .dst_cpu = target_cpu,
5595 .dst_rq = target_rq,
5596 .src_cpu = busiest_rq->cpu,
5597 .src_rq = busiest_rq,
5601 schedstat_inc(sd, alb_count);
5603 if (move_one_task(&env))
5604 schedstat_inc(sd, alb_pushed);
5606 schedstat_inc(sd, alb_failed);
5609 double_unlock_balance(busiest_rq, target_rq);
5611 busiest_rq->active_balance = 0;
5612 raw_spin_unlock_irq(&busiest_rq->lock);
5616 #ifdef CONFIG_NO_HZ_COMMON
5618 * idle load balancing details
5619 * - When one of the busy CPUs notice that there may be an idle rebalancing
5620 * needed, they will kick the idle load balancer, which then does idle
5621 * load balancing for all the idle CPUs.
5624 cpumask_var_t idle_cpus_mask;
5626 unsigned long next_balance; /* in jiffy units */
5627 } nohz ____cacheline_aligned;
5629 static inline int find_new_ilb(int call_cpu)
5631 int ilb = cpumask_first(nohz.idle_cpus_mask);
5633 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5640 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5641 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5642 * CPU (if there is one).
5644 static void nohz_balancer_kick(int cpu)
5648 nohz.next_balance++;
5650 ilb_cpu = find_new_ilb(cpu);
5652 if (ilb_cpu >= nr_cpu_ids)
5655 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5658 * Use smp_send_reschedule() instead of resched_cpu().
5659 * This way we generate a sched IPI on the target cpu which
5660 * is idle. And the softirq performing nohz idle load balance
5661 * will be run before returning from the IPI.
5663 smp_send_reschedule(ilb_cpu);
5667 static inline void nohz_balance_exit_idle(int cpu)
5669 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5670 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5671 atomic_dec(&nohz.nr_cpus);
5672 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5676 static inline void set_cpu_sd_state_busy(void)
5678 struct sched_domain *sd;
5681 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5683 if (!sd || !sd->nohz_idle)
5687 for (; sd; sd = sd->parent)
5688 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5693 void set_cpu_sd_state_idle(void)
5695 struct sched_domain *sd;
5698 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5700 if (!sd || sd->nohz_idle)
5704 for (; sd; sd = sd->parent)
5705 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5711 * This routine will record that the cpu is going idle with tick stopped.
5712 * This info will be used in performing idle load balancing in the future.
5714 void nohz_balance_enter_idle(int cpu)
5717 * If this cpu is going down, then nothing needs to be done.
5719 if (!cpu_active(cpu))
5722 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5725 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5726 atomic_inc(&nohz.nr_cpus);
5727 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5730 static int sched_ilb_notifier(struct notifier_block *nfb,
5731 unsigned long action, void *hcpu)
5733 switch (action & ~CPU_TASKS_FROZEN) {
5735 nohz_balance_exit_idle(smp_processor_id());
5743 static DEFINE_SPINLOCK(balancing);
5746 * Scale the max load_balance interval with the number of CPUs in the system.
5747 * This trades load-balance latency on larger machines for less cross talk.
5749 void update_max_interval(void)
5751 max_load_balance_interval = HZ*num_online_cpus()/10;
5755 * It checks each scheduling domain to see if it is due to be balanced,
5756 * and initiates a balancing operation if so.
5758 * Balancing parameters are set up in init_sched_domains.
5760 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5762 int continue_balancing = 1;
5763 struct rq *rq = cpu_rq(cpu);
5764 unsigned long interval;
5765 struct sched_domain *sd;
5766 /* Earliest time when we have to do rebalance again */
5767 unsigned long next_balance = jiffies + 60*HZ;
5768 int update_next_balance = 0;
5769 int need_serialize, need_decay = 0;
5772 update_blocked_averages(cpu);
5775 for_each_domain(cpu, sd) {
5777 * Decay the newidle max times here because this is a regular
5778 * visit to all the domains. Decay ~1% per second.
5780 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5781 sd->max_newidle_lb_cost =
5782 (sd->max_newidle_lb_cost * 253) / 256;
5783 sd->next_decay_max_lb_cost = jiffies + HZ;
5786 max_cost += sd->max_newidle_lb_cost;
5788 if (!(sd->flags & SD_LOAD_BALANCE))
5792 * Stop the load balance at this level. There is another
5793 * CPU in our sched group which is doing load balancing more
5796 if (!continue_balancing) {
5802 interval = sd->balance_interval;
5803 if (idle != CPU_IDLE)
5804 interval *= sd->busy_factor;
5806 /* scale ms to jiffies */
5807 interval = msecs_to_jiffies(interval);
5808 interval = clamp(interval, 1UL, max_load_balance_interval);
5810 need_serialize = sd->flags & SD_SERIALIZE;
5812 if (need_serialize) {
5813 if (!spin_trylock(&balancing))
5817 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5818 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5820 * The LBF_DST_PINNED logic could have changed
5821 * env->dst_cpu, so we can't know our idle
5822 * state even if we migrated tasks. Update it.
5824 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5826 sd->last_balance = jiffies;
5829 spin_unlock(&balancing);
5831 if (time_after(next_balance, sd->last_balance + interval)) {
5832 next_balance = sd->last_balance + interval;
5833 update_next_balance = 1;
5838 * Ensure the rq-wide value also decays but keep it at a
5839 * reasonable floor to avoid funnies with rq->avg_idle.
5841 rq->max_idle_balance_cost =
5842 max((u64)sysctl_sched_migration_cost, max_cost);
5847 * next_balance will be updated only when there is a need.
5848 * When the cpu is attached to null domain for ex, it will not be
5851 if (likely(update_next_balance))
5852 rq->next_balance = next_balance;
5855 #ifdef CONFIG_NO_HZ_COMMON
5857 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5858 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5860 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5862 struct rq *this_rq = cpu_rq(this_cpu);
5866 if (idle != CPU_IDLE ||
5867 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5870 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5871 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5875 * If this cpu gets work to do, stop the load balancing
5876 * work being done for other cpus. Next load
5877 * balancing owner will pick it up.
5882 rq = cpu_rq(balance_cpu);
5884 raw_spin_lock_irq(&rq->lock);
5885 update_rq_clock(rq);
5886 update_idle_cpu_load(rq);
5887 raw_spin_unlock_irq(&rq->lock);
5889 rebalance_domains(balance_cpu, CPU_IDLE);
5891 if (time_after(this_rq->next_balance, rq->next_balance))
5892 this_rq->next_balance = rq->next_balance;
5894 nohz.next_balance = this_rq->next_balance;
5896 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5900 * Current heuristic for kicking the idle load balancer in the presence
5901 * of an idle cpu is the system.
5902 * - This rq has more than one task.
5903 * - At any scheduler domain level, this cpu's scheduler group has multiple
5904 * busy cpu's exceeding the group's power.
5905 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5906 * domain span are idle.
5908 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5910 unsigned long now = jiffies;
5911 struct sched_domain *sd;
5913 if (unlikely(idle_cpu(cpu)))
5917 * We may be recently in ticked or tickless idle mode. At the first
5918 * busy tick after returning from idle, we will update the busy stats.
5920 set_cpu_sd_state_busy();
5921 nohz_balance_exit_idle(cpu);
5924 * None are in tickless mode and hence no need for NOHZ idle load
5927 if (likely(!atomic_read(&nohz.nr_cpus)))
5930 if (time_before(now, nohz.next_balance))
5933 if (rq->nr_running >= 2)
5937 for_each_domain(cpu, sd) {
5938 struct sched_group *sg = sd->groups;
5939 struct sched_group_power *sgp = sg->sgp;
5940 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5942 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5943 goto need_kick_unlock;
5945 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5946 && (cpumask_first_and(nohz.idle_cpus_mask,
5947 sched_domain_span(sd)) < cpu))
5948 goto need_kick_unlock;
5950 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5962 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5966 * run_rebalance_domains is triggered when needed from the scheduler tick.
5967 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5969 static void run_rebalance_domains(struct softirq_action *h)
5971 int this_cpu = smp_processor_id();
5972 struct rq *this_rq = cpu_rq(this_cpu);
5973 enum cpu_idle_type idle = this_rq->idle_balance ?
5974 CPU_IDLE : CPU_NOT_IDLE;
5976 rebalance_domains(this_cpu, idle);
5979 * If this cpu has a pending nohz_balance_kick, then do the
5980 * balancing on behalf of the other idle cpus whose ticks are
5983 nohz_idle_balance(this_cpu, idle);
5986 static inline int on_null_domain(int cpu)
5988 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5992 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5994 void trigger_load_balance(struct rq *rq, int cpu)
5996 /* Don't need to rebalance while attached to NULL domain */
5997 if (time_after_eq(jiffies, rq->next_balance) &&
5998 likely(!on_null_domain(cpu)))
5999 raise_softirq(SCHED_SOFTIRQ);
6000 #ifdef CONFIG_NO_HZ_COMMON
6001 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6002 nohz_balancer_kick(cpu);
6006 static void rq_online_fair(struct rq *rq)
6011 static void rq_offline_fair(struct rq *rq)
6015 /* Ensure any throttled groups are reachable by pick_next_task */
6016 unthrottle_offline_cfs_rqs(rq);
6019 #endif /* CONFIG_SMP */
6022 * scheduler tick hitting a task of our scheduling class:
6024 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6026 struct cfs_rq *cfs_rq;
6027 struct sched_entity *se = &curr->se;
6029 for_each_sched_entity(se) {
6030 cfs_rq = cfs_rq_of(se);
6031 entity_tick(cfs_rq, se, queued);
6034 if (numabalancing_enabled)
6035 task_tick_numa(rq, curr);
6037 update_rq_runnable_avg(rq, 1);
6041 * called on fork with the child task as argument from the parent's context
6042 * - child not yet on the tasklist
6043 * - preemption disabled
6045 static void task_fork_fair(struct task_struct *p)
6047 struct cfs_rq *cfs_rq;
6048 struct sched_entity *se = &p->se, *curr;
6049 int this_cpu = smp_processor_id();
6050 struct rq *rq = this_rq();
6051 unsigned long flags;
6053 raw_spin_lock_irqsave(&rq->lock, flags);
6055 update_rq_clock(rq);
6057 cfs_rq = task_cfs_rq(current);
6058 curr = cfs_rq->curr;
6061 * Not only the cpu but also the task_group of the parent might have
6062 * been changed after parent->se.parent,cfs_rq were copied to
6063 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6064 * of child point to valid ones.
6067 __set_task_cpu(p, this_cpu);
6070 update_curr(cfs_rq);
6073 se->vruntime = curr->vruntime;
6074 place_entity(cfs_rq, se, 1);
6076 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6078 * Upon rescheduling, sched_class::put_prev_task() will place
6079 * 'current' within the tree based on its new key value.
6081 swap(curr->vruntime, se->vruntime);
6082 resched_task(rq->curr);
6085 se->vruntime -= cfs_rq->min_vruntime;
6087 raw_spin_unlock_irqrestore(&rq->lock, flags);
6091 * Priority of the task has changed. Check to see if we preempt
6095 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6101 * Reschedule if we are currently running on this runqueue and
6102 * our priority decreased, or if we are not currently running on
6103 * this runqueue and our priority is higher than the current's
6105 if (rq->curr == p) {
6106 if (p->prio > oldprio)
6107 resched_task(rq->curr);
6109 check_preempt_curr(rq, p, 0);
6112 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6114 struct sched_entity *se = &p->se;
6115 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6118 * Ensure the task's vruntime is normalized, so that when its
6119 * switched back to the fair class the enqueue_entity(.flags=0) will
6120 * do the right thing.
6122 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6123 * have normalized the vruntime, if it was !on_rq, then only when
6124 * the task is sleeping will it still have non-normalized vruntime.
6126 if (!se->on_rq && p->state != TASK_RUNNING) {
6128 * Fix up our vruntime so that the current sleep doesn't
6129 * cause 'unlimited' sleep bonus.
6131 place_entity(cfs_rq, se, 0);
6132 se->vruntime -= cfs_rq->min_vruntime;
6137 * Remove our load from contribution when we leave sched_fair
6138 * and ensure we don't carry in an old decay_count if we
6141 if (se->avg.decay_count) {
6142 __synchronize_entity_decay(se);
6143 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6149 * We switched to the sched_fair class.
6151 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6157 * We were most likely switched from sched_rt, so
6158 * kick off the schedule if running, otherwise just see
6159 * if we can still preempt the current task.
6162 resched_task(rq->curr);
6164 check_preempt_curr(rq, p, 0);
6167 /* Account for a task changing its policy or group.
6169 * This routine is mostly called to set cfs_rq->curr field when a task
6170 * migrates between groups/classes.
6172 static void set_curr_task_fair(struct rq *rq)
6174 struct sched_entity *se = &rq->curr->se;
6176 for_each_sched_entity(se) {
6177 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6179 set_next_entity(cfs_rq, se);
6180 /* ensure bandwidth has been allocated on our new cfs_rq */
6181 account_cfs_rq_runtime(cfs_rq, 0);
6185 void init_cfs_rq(struct cfs_rq *cfs_rq)
6187 cfs_rq->tasks_timeline = RB_ROOT;
6188 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6189 #ifndef CONFIG_64BIT
6190 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6193 atomic64_set(&cfs_rq->decay_counter, 1);
6194 atomic_long_set(&cfs_rq->removed_load, 0);
6198 #ifdef CONFIG_FAIR_GROUP_SCHED
6199 static void task_move_group_fair(struct task_struct *p, int on_rq)
6201 struct cfs_rq *cfs_rq;
6203 * If the task was not on the rq at the time of this cgroup movement
6204 * it must have been asleep, sleeping tasks keep their ->vruntime
6205 * absolute on their old rq until wakeup (needed for the fair sleeper
6206 * bonus in place_entity()).
6208 * If it was on the rq, we've just 'preempted' it, which does convert
6209 * ->vruntime to a relative base.
6211 * Make sure both cases convert their relative position when migrating
6212 * to another cgroup's rq. This does somewhat interfere with the
6213 * fair sleeper stuff for the first placement, but who cares.
6216 * When !on_rq, vruntime of the task has usually NOT been normalized.
6217 * But there are some cases where it has already been normalized:
6219 * - Moving a forked child which is waiting for being woken up by
6220 * wake_up_new_task().
6221 * - Moving a task which has been woken up by try_to_wake_up() and
6222 * waiting for actually being woken up by sched_ttwu_pending().
6224 * To prevent boost or penalty in the new cfs_rq caused by delta
6225 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6227 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6231 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6232 set_task_rq(p, task_cpu(p));
6234 cfs_rq = cfs_rq_of(&p->se);
6235 p->se.vruntime += cfs_rq->min_vruntime;
6238 * migrate_task_rq_fair() will have removed our previous
6239 * contribution, but we must synchronize for ongoing future
6242 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6243 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6248 void free_fair_sched_group(struct task_group *tg)
6252 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6254 for_each_possible_cpu(i) {
6256 kfree(tg->cfs_rq[i]);
6265 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6267 struct cfs_rq *cfs_rq;
6268 struct sched_entity *se;
6271 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6274 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6278 tg->shares = NICE_0_LOAD;
6280 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6282 for_each_possible_cpu(i) {
6283 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6284 GFP_KERNEL, cpu_to_node(i));
6288 se = kzalloc_node(sizeof(struct sched_entity),
6289 GFP_KERNEL, cpu_to_node(i));
6293 init_cfs_rq(cfs_rq);
6294 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6305 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6307 struct rq *rq = cpu_rq(cpu);
6308 unsigned long flags;
6311 * Only empty task groups can be destroyed; so we can speculatively
6312 * check on_list without danger of it being re-added.
6314 if (!tg->cfs_rq[cpu]->on_list)
6317 raw_spin_lock_irqsave(&rq->lock, flags);
6318 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6319 raw_spin_unlock_irqrestore(&rq->lock, flags);
6322 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6323 struct sched_entity *se, int cpu,
6324 struct sched_entity *parent)
6326 struct rq *rq = cpu_rq(cpu);
6330 init_cfs_rq_runtime(cfs_rq);
6332 tg->cfs_rq[cpu] = cfs_rq;
6335 /* se could be NULL for root_task_group */
6340 se->cfs_rq = &rq->cfs;
6342 se->cfs_rq = parent->my_q;
6345 update_load_set(&se->load, 0);
6346 se->parent = parent;
6349 static DEFINE_MUTEX(shares_mutex);
6351 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6354 unsigned long flags;
6357 * We can't change the weight of the root cgroup.
6362 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6364 mutex_lock(&shares_mutex);
6365 if (tg->shares == shares)
6368 tg->shares = shares;
6369 for_each_possible_cpu(i) {
6370 struct rq *rq = cpu_rq(i);
6371 struct sched_entity *se;
6374 /* Propagate contribution to hierarchy */
6375 raw_spin_lock_irqsave(&rq->lock, flags);
6377 /* Possible calls to update_curr() need rq clock */
6378 update_rq_clock(rq);
6379 for_each_sched_entity(se)
6380 update_cfs_shares(group_cfs_rq(se));
6381 raw_spin_unlock_irqrestore(&rq->lock, flags);
6385 mutex_unlock(&shares_mutex);
6388 #else /* CONFIG_FAIR_GROUP_SCHED */
6390 void free_fair_sched_group(struct task_group *tg) { }
6392 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6397 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6399 #endif /* CONFIG_FAIR_GROUP_SCHED */
6402 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6404 struct sched_entity *se = &task->se;
6405 unsigned int rr_interval = 0;
6408 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6411 if (rq->cfs.load.weight)
6412 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6418 * All the scheduling class methods:
6420 const struct sched_class fair_sched_class = {
6421 .next = &idle_sched_class,
6422 .enqueue_task = enqueue_task_fair,
6423 .dequeue_task = dequeue_task_fair,
6424 .yield_task = yield_task_fair,
6425 .yield_to_task = yield_to_task_fair,
6427 .check_preempt_curr = check_preempt_wakeup,
6429 .pick_next_task = pick_next_task_fair,
6430 .put_prev_task = put_prev_task_fair,
6433 .select_task_rq = select_task_rq_fair,
6434 .migrate_task_rq = migrate_task_rq_fair,
6436 .rq_online = rq_online_fair,
6437 .rq_offline = rq_offline_fair,
6439 .task_waking = task_waking_fair,
6442 .set_curr_task = set_curr_task_fair,
6443 .task_tick = task_tick_fair,
6444 .task_fork = task_fork_fair,
6446 .prio_changed = prio_changed_fair,
6447 .switched_from = switched_from_fair,
6448 .switched_to = switched_to_fair,
6450 .get_rr_interval = get_rr_interval_fair,
6452 #ifdef CONFIG_FAIR_GROUP_SCHED
6453 .task_move_group = task_move_group_fair,
6457 #ifdef CONFIG_SCHED_DEBUG
6458 void print_cfs_stats(struct seq_file *m, int cpu)
6460 struct cfs_rq *cfs_rq;
6463 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6464 print_cfs_rq(m, cpu, cfs_rq);
6469 __init void init_sched_fair_class(void)
6472 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6474 #ifdef CONFIG_NO_HZ_COMMON
6475 nohz.next_balance = jiffies;
6476 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6477 cpu_notifier(sched_ilb_notifier, 0);