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);
881 * Once a preferred node is selected the scheduler balancer will prefer moving
882 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
883 * scans. This will give the process the chance to accumulate more faults on
884 * the preferred node but still allow the scheduler to move the task again if
885 * the nodes CPUs are overloaded.
887 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 3;
889 static void task_numa_placement(struct task_struct *p)
891 int seq, nid, max_nid = -1;
892 unsigned long max_faults = 0;
894 if (!p->mm) /* for example, ksmd faulting in a user's mm */
896 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
897 if (p->numa_scan_seq == seq)
899 p->numa_scan_seq = seq;
900 p->numa_migrate_seq++;
901 p->numa_scan_period_max = task_scan_max(p);
903 /* Find the node with the highest number of faults */
904 for_each_online_node(nid) {
905 unsigned long faults;
907 /* Decay existing window and copy faults since last scan */
908 p->numa_faults[nid] >>= 1;
909 p->numa_faults[nid] += p->numa_faults_buffer[nid];
910 p->numa_faults_buffer[nid] = 0;
912 faults = p->numa_faults[nid];
913 if (faults > max_faults) {
919 /* Update the tasks preferred node if necessary */
920 if (max_faults && max_nid != p->numa_preferred_nid) {
921 p->numa_preferred_nid = max_nid;
922 p->numa_migrate_seq = 0;
927 * Got a PROT_NONE fault for a page on @node.
929 void task_numa_fault(int node, int pages, bool migrated)
931 struct task_struct *p = current;
933 if (!numabalancing_enabled)
936 /* Allocate buffer to track faults on a per-node basis */
937 if (unlikely(!p->numa_faults)) {
938 int size = sizeof(*p->numa_faults) * nr_node_ids;
940 /* numa_faults and numa_faults_buffer share the allocation */
941 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
945 BUG_ON(p->numa_faults_buffer);
946 p->numa_faults_buffer = p->numa_faults + nr_node_ids;
950 * If pages are properly placed (did not migrate) then scan slower.
951 * This is reset periodically in case of phase changes
954 /* Initialise if necessary */
955 if (!p->numa_scan_period_max)
956 p->numa_scan_period_max = task_scan_max(p);
958 p->numa_scan_period = min(p->numa_scan_period_max,
959 p->numa_scan_period + 10);
962 task_numa_placement(p);
964 p->numa_faults_buffer[node] += pages;
967 static void reset_ptenuma_scan(struct task_struct *p)
969 ACCESS_ONCE(p->mm->numa_scan_seq)++;
970 p->mm->numa_scan_offset = 0;
974 * The expensive part of numa migration is done from task_work context.
975 * Triggered from task_tick_numa().
977 void task_numa_work(struct callback_head *work)
979 unsigned long migrate, next_scan, now = jiffies;
980 struct task_struct *p = current;
981 struct mm_struct *mm = p->mm;
982 struct vm_area_struct *vma;
983 unsigned long start, end;
984 unsigned long nr_pte_updates = 0;
987 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
989 work->next = work; /* protect against double add */
991 * Who cares about NUMA placement when they're dying.
993 * NOTE: make sure not to dereference p->mm before this check,
994 * exit_task_work() happens _after_ exit_mm() so we could be called
995 * without p->mm even though we still had it when we enqueued this
998 if (p->flags & PF_EXITING)
1001 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1002 mm->numa_next_scan = now +
1003 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1004 mm->numa_next_reset = now +
1005 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1009 * Reset the scan period if enough time has gone by. Objective is that
1010 * scanning will be reduced if pages are properly placed. As tasks
1011 * can enter different phases this needs to be re-examined. Lacking
1012 * proper tracking of reference behaviour, this blunt hammer is used.
1014 migrate = mm->numa_next_reset;
1015 if (time_after(now, migrate)) {
1016 p->numa_scan_period = task_scan_min(p);
1017 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1018 xchg(&mm->numa_next_reset, next_scan);
1022 * Enforce maximal scan/migration frequency..
1024 migrate = mm->numa_next_scan;
1025 if (time_before(now, migrate))
1028 if (p->numa_scan_period == 0) {
1029 p->numa_scan_period_max = task_scan_max(p);
1030 p->numa_scan_period = task_scan_min(p);
1033 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1034 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1038 * Delay this task enough that another task of this mm will likely win
1039 * the next time around.
1041 p->node_stamp += 2 * TICK_NSEC;
1043 start = mm->numa_scan_offset;
1044 pages = sysctl_numa_balancing_scan_size;
1045 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1049 down_read(&mm->mmap_sem);
1050 vma = find_vma(mm, start);
1052 reset_ptenuma_scan(p);
1056 for (; vma; vma = vma->vm_next) {
1057 if (!vma_migratable(vma))
1060 /* Skip small VMAs. They are not likely to be of relevance */
1061 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1065 start = max(start, vma->vm_start);
1066 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1067 end = min(end, vma->vm_end);
1068 nr_pte_updates += change_prot_numa(vma, start, end);
1071 * Scan sysctl_numa_balancing_scan_size but ensure that
1072 * at least one PTE is updated so that unused virtual
1073 * address space is quickly skipped.
1076 pages -= (end - start) >> PAGE_SHIFT;
1081 } while (end != vma->vm_end);
1086 * If the whole process was scanned without updates then no NUMA
1087 * hinting faults are being recorded and scan rate should be lower.
1089 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1090 p->numa_scan_period = min(p->numa_scan_period_max,
1091 p->numa_scan_period << 1);
1093 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1094 mm->numa_next_scan = next_scan;
1098 * It is possible to reach the end of the VMA list but the last few
1099 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1100 * would find the !migratable VMA on the next scan but not reset the
1101 * scanner to the start so check it now.
1104 mm->numa_scan_offset = start;
1106 reset_ptenuma_scan(p);
1107 up_read(&mm->mmap_sem);
1111 * Drive the periodic memory faults..
1113 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1115 struct callback_head *work = &curr->numa_work;
1119 * We don't care about NUMA placement if we don't have memory.
1121 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1125 * Using runtime rather than walltime has the dual advantage that
1126 * we (mostly) drive the selection from busy threads and that the
1127 * task needs to have done some actual work before we bother with
1130 now = curr->se.sum_exec_runtime;
1131 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1133 if (now - curr->node_stamp > period) {
1134 if (!curr->node_stamp)
1135 curr->numa_scan_period = task_scan_min(curr);
1136 curr->node_stamp += period;
1138 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1139 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1140 task_work_add(curr, work, true);
1145 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1148 #endif /* CONFIG_NUMA_BALANCING */
1151 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1153 update_load_add(&cfs_rq->load, se->load.weight);
1154 if (!parent_entity(se))
1155 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1157 if (entity_is_task(se))
1158 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1160 cfs_rq->nr_running++;
1164 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1166 update_load_sub(&cfs_rq->load, se->load.weight);
1167 if (!parent_entity(se))
1168 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1169 if (entity_is_task(se))
1170 list_del_init(&se->group_node);
1171 cfs_rq->nr_running--;
1174 #ifdef CONFIG_FAIR_GROUP_SCHED
1176 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1181 * Use this CPU's actual weight instead of the last load_contribution
1182 * to gain a more accurate current total weight. See
1183 * update_cfs_rq_load_contribution().
1185 tg_weight = atomic_long_read(&tg->load_avg);
1186 tg_weight -= cfs_rq->tg_load_contrib;
1187 tg_weight += cfs_rq->load.weight;
1192 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1194 long tg_weight, load, shares;
1196 tg_weight = calc_tg_weight(tg, cfs_rq);
1197 load = cfs_rq->load.weight;
1199 shares = (tg->shares * load);
1201 shares /= tg_weight;
1203 if (shares < MIN_SHARES)
1204 shares = MIN_SHARES;
1205 if (shares > tg->shares)
1206 shares = tg->shares;
1210 # else /* CONFIG_SMP */
1211 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1215 # endif /* CONFIG_SMP */
1216 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1217 unsigned long weight)
1220 /* commit outstanding execution time */
1221 if (cfs_rq->curr == se)
1222 update_curr(cfs_rq);
1223 account_entity_dequeue(cfs_rq, se);
1226 update_load_set(&se->load, weight);
1229 account_entity_enqueue(cfs_rq, se);
1232 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1234 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1236 struct task_group *tg;
1237 struct sched_entity *se;
1241 se = tg->se[cpu_of(rq_of(cfs_rq))];
1242 if (!se || throttled_hierarchy(cfs_rq))
1245 if (likely(se->load.weight == tg->shares))
1248 shares = calc_cfs_shares(cfs_rq, tg);
1250 reweight_entity(cfs_rq_of(se), se, shares);
1252 #else /* CONFIG_FAIR_GROUP_SCHED */
1253 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1256 #endif /* CONFIG_FAIR_GROUP_SCHED */
1260 * We choose a half-life close to 1 scheduling period.
1261 * Note: The tables below are dependent on this value.
1263 #define LOAD_AVG_PERIOD 32
1264 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1265 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1267 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1268 static const u32 runnable_avg_yN_inv[] = {
1269 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1270 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1271 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1272 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1273 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1274 0x85aac367, 0x82cd8698,
1278 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1279 * over-estimates when re-combining.
1281 static const u32 runnable_avg_yN_sum[] = {
1282 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1283 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1284 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1289 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1291 static __always_inline u64 decay_load(u64 val, u64 n)
1293 unsigned int local_n;
1297 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1300 /* after bounds checking we can collapse to 32-bit */
1304 * As y^PERIOD = 1/2, we can combine
1305 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1306 * With a look-up table which covers k^n (n<PERIOD)
1308 * To achieve constant time decay_load.
1310 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1311 val >>= local_n / LOAD_AVG_PERIOD;
1312 local_n %= LOAD_AVG_PERIOD;
1315 val *= runnable_avg_yN_inv[local_n];
1316 /* We don't use SRR here since we always want to round down. */
1321 * For updates fully spanning n periods, the contribution to runnable
1322 * average will be: \Sum 1024*y^n
1324 * We can compute this reasonably efficiently by combining:
1325 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1327 static u32 __compute_runnable_contrib(u64 n)
1331 if (likely(n <= LOAD_AVG_PERIOD))
1332 return runnable_avg_yN_sum[n];
1333 else if (unlikely(n >= LOAD_AVG_MAX_N))
1334 return LOAD_AVG_MAX;
1336 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1338 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1339 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1341 n -= LOAD_AVG_PERIOD;
1342 } while (n > LOAD_AVG_PERIOD);
1344 contrib = decay_load(contrib, n);
1345 return contrib + runnable_avg_yN_sum[n];
1349 * We can represent the historical contribution to runnable average as the
1350 * coefficients of a geometric series. To do this we sub-divide our runnable
1351 * history into segments of approximately 1ms (1024us); label the segment that
1352 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1354 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1356 * (now) (~1ms ago) (~2ms ago)
1358 * Let u_i denote the fraction of p_i that the entity was runnable.
1360 * We then designate the fractions u_i as our co-efficients, yielding the
1361 * following representation of historical load:
1362 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1364 * We choose y based on the with of a reasonably scheduling period, fixing:
1367 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1368 * approximately half as much as the contribution to load within the last ms
1371 * When a period "rolls over" and we have new u_0`, multiplying the previous
1372 * sum again by y is sufficient to update:
1373 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1374 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1376 static __always_inline int __update_entity_runnable_avg(u64 now,
1377 struct sched_avg *sa,
1381 u32 runnable_contrib;
1382 int delta_w, decayed = 0;
1384 delta = now - sa->last_runnable_update;
1386 * This should only happen when time goes backwards, which it
1387 * unfortunately does during sched clock init when we swap over to TSC.
1389 if ((s64)delta < 0) {
1390 sa->last_runnable_update = now;
1395 * Use 1024ns as the unit of measurement since it's a reasonable
1396 * approximation of 1us and fast to compute.
1401 sa->last_runnable_update = now;
1403 /* delta_w is the amount already accumulated against our next period */
1404 delta_w = sa->runnable_avg_period % 1024;
1405 if (delta + delta_w >= 1024) {
1406 /* period roll-over */
1410 * Now that we know we're crossing a period boundary, figure
1411 * out how much from delta we need to complete the current
1412 * period and accrue it.
1414 delta_w = 1024 - delta_w;
1416 sa->runnable_avg_sum += delta_w;
1417 sa->runnable_avg_period += delta_w;
1421 /* Figure out how many additional periods this update spans */
1422 periods = delta / 1024;
1425 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1427 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1430 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1431 runnable_contrib = __compute_runnable_contrib(periods);
1433 sa->runnable_avg_sum += runnable_contrib;
1434 sa->runnable_avg_period += runnable_contrib;
1437 /* Remainder of delta accrued against u_0` */
1439 sa->runnable_avg_sum += delta;
1440 sa->runnable_avg_period += delta;
1445 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1446 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1448 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1449 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1451 decays -= se->avg.decay_count;
1455 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1456 se->avg.decay_count = 0;
1461 #ifdef CONFIG_FAIR_GROUP_SCHED
1462 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1465 struct task_group *tg = cfs_rq->tg;
1468 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1469 tg_contrib -= cfs_rq->tg_load_contrib;
1471 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1472 atomic_long_add(tg_contrib, &tg->load_avg);
1473 cfs_rq->tg_load_contrib += tg_contrib;
1478 * Aggregate cfs_rq runnable averages into an equivalent task_group
1479 * representation for computing load contributions.
1481 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1482 struct cfs_rq *cfs_rq)
1484 struct task_group *tg = cfs_rq->tg;
1487 /* The fraction of a cpu used by this cfs_rq */
1488 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1489 sa->runnable_avg_period + 1);
1490 contrib -= cfs_rq->tg_runnable_contrib;
1492 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1493 atomic_add(contrib, &tg->runnable_avg);
1494 cfs_rq->tg_runnable_contrib += contrib;
1498 static inline void __update_group_entity_contrib(struct sched_entity *se)
1500 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1501 struct task_group *tg = cfs_rq->tg;
1506 contrib = cfs_rq->tg_load_contrib * tg->shares;
1507 se->avg.load_avg_contrib = div_u64(contrib,
1508 atomic_long_read(&tg->load_avg) + 1);
1511 * For group entities we need to compute a correction term in the case
1512 * that they are consuming <1 cpu so that we would contribute the same
1513 * load as a task of equal weight.
1515 * Explicitly co-ordinating this measurement would be expensive, but
1516 * fortunately the sum of each cpus contribution forms a usable
1517 * lower-bound on the true value.
1519 * Consider the aggregate of 2 contributions. Either they are disjoint
1520 * (and the sum represents true value) or they are disjoint and we are
1521 * understating by the aggregate of their overlap.
1523 * Extending this to N cpus, for a given overlap, the maximum amount we
1524 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1525 * cpus that overlap for this interval and w_i is the interval width.
1527 * On a small machine; the first term is well-bounded which bounds the
1528 * total error since w_i is a subset of the period. Whereas on a
1529 * larger machine, while this first term can be larger, if w_i is the
1530 * of consequential size guaranteed to see n_i*w_i quickly converge to
1531 * our upper bound of 1-cpu.
1533 runnable_avg = atomic_read(&tg->runnable_avg);
1534 if (runnable_avg < NICE_0_LOAD) {
1535 se->avg.load_avg_contrib *= runnable_avg;
1536 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1540 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1541 int force_update) {}
1542 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1543 struct cfs_rq *cfs_rq) {}
1544 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1547 static inline void __update_task_entity_contrib(struct sched_entity *se)
1551 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1552 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1553 contrib /= (se->avg.runnable_avg_period + 1);
1554 se->avg.load_avg_contrib = scale_load(contrib);
1557 /* Compute the current contribution to load_avg by se, return any delta */
1558 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1560 long old_contrib = se->avg.load_avg_contrib;
1562 if (entity_is_task(se)) {
1563 __update_task_entity_contrib(se);
1565 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1566 __update_group_entity_contrib(se);
1569 return se->avg.load_avg_contrib - old_contrib;
1572 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1575 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1576 cfs_rq->blocked_load_avg -= load_contrib;
1578 cfs_rq->blocked_load_avg = 0;
1581 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1583 /* Update a sched_entity's runnable average */
1584 static inline void update_entity_load_avg(struct sched_entity *se,
1587 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1592 * For a group entity we need to use their owned cfs_rq_clock_task() in
1593 * case they are the parent of a throttled hierarchy.
1595 if (entity_is_task(se))
1596 now = cfs_rq_clock_task(cfs_rq);
1598 now = cfs_rq_clock_task(group_cfs_rq(se));
1600 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1603 contrib_delta = __update_entity_load_avg_contrib(se);
1609 cfs_rq->runnable_load_avg += contrib_delta;
1611 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1615 * Decay the load contributed by all blocked children and account this so that
1616 * their contribution may appropriately discounted when they wake up.
1618 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1620 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1623 decays = now - cfs_rq->last_decay;
1624 if (!decays && !force_update)
1627 if (atomic_long_read(&cfs_rq->removed_load)) {
1628 unsigned long removed_load;
1629 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1630 subtract_blocked_load_contrib(cfs_rq, removed_load);
1634 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1636 atomic64_add(decays, &cfs_rq->decay_counter);
1637 cfs_rq->last_decay = now;
1640 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1643 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1645 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1646 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1649 /* Add the load generated by se into cfs_rq's child load-average */
1650 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1651 struct sched_entity *se,
1655 * We track migrations using entity decay_count <= 0, on a wake-up
1656 * migration we use a negative decay count to track the remote decays
1657 * accumulated while sleeping.
1659 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1660 * are seen by enqueue_entity_load_avg() as a migration with an already
1661 * constructed load_avg_contrib.
1663 if (unlikely(se->avg.decay_count <= 0)) {
1664 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1665 if (se->avg.decay_count) {
1667 * In a wake-up migration we have to approximate the
1668 * time sleeping. This is because we can't synchronize
1669 * clock_task between the two cpus, and it is not
1670 * guaranteed to be read-safe. Instead, we can
1671 * approximate this using our carried decays, which are
1672 * explicitly atomically readable.
1674 se->avg.last_runnable_update -= (-se->avg.decay_count)
1676 update_entity_load_avg(se, 0);
1677 /* Indicate that we're now synchronized and on-rq */
1678 se->avg.decay_count = 0;
1683 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1684 * would have made count negative); we must be careful to avoid
1685 * double-accounting blocked time after synchronizing decays.
1687 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1691 /* migrated tasks did not contribute to our blocked load */
1693 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1694 update_entity_load_avg(se, 0);
1697 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1698 /* we force update consideration on load-balancer moves */
1699 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1703 * Remove se's load from this cfs_rq child load-average, if the entity is
1704 * transitioning to a blocked state we track its projected decay using
1707 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1708 struct sched_entity *se,
1711 update_entity_load_avg(se, 1);
1712 /* we force update consideration on load-balancer moves */
1713 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1715 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1717 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1718 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1719 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1723 * Update the rq's load with the elapsed running time before entering
1724 * idle. if the last scheduled task is not a CFS task, idle_enter will
1725 * be the only way to update the runnable statistic.
1727 void idle_enter_fair(struct rq *this_rq)
1729 update_rq_runnable_avg(this_rq, 1);
1733 * Update the rq's load with the elapsed idle time before a task is
1734 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1735 * be the only way to update the runnable statistic.
1737 void idle_exit_fair(struct rq *this_rq)
1739 update_rq_runnable_avg(this_rq, 0);
1743 static inline void update_entity_load_avg(struct sched_entity *se,
1744 int update_cfs_rq) {}
1745 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1746 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1747 struct sched_entity *se,
1749 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1750 struct sched_entity *se,
1752 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1753 int force_update) {}
1756 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1758 #ifdef CONFIG_SCHEDSTATS
1759 struct task_struct *tsk = NULL;
1761 if (entity_is_task(se))
1764 if (se->statistics.sleep_start) {
1765 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1770 if (unlikely(delta > se->statistics.sleep_max))
1771 se->statistics.sleep_max = delta;
1773 se->statistics.sleep_start = 0;
1774 se->statistics.sum_sleep_runtime += delta;
1777 account_scheduler_latency(tsk, delta >> 10, 1);
1778 trace_sched_stat_sleep(tsk, delta);
1781 if (se->statistics.block_start) {
1782 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1787 if (unlikely(delta > se->statistics.block_max))
1788 se->statistics.block_max = delta;
1790 se->statistics.block_start = 0;
1791 se->statistics.sum_sleep_runtime += delta;
1794 if (tsk->in_iowait) {
1795 se->statistics.iowait_sum += delta;
1796 se->statistics.iowait_count++;
1797 trace_sched_stat_iowait(tsk, delta);
1800 trace_sched_stat_blocked(tsk, delta);
1803 * Blocking time is in units of nanosecs, so shift by
1804 * 20 to get a milliseconds-range estimation of the
1805 * amount of time that the task spent sleeping:
1807 if (unlikely(prof_on == SLEEP_PROFILING)) {
1808 profile_hits(SLEEP_PROFILING,
1809 (void *)get_wchan(tsk),
1812 account_scheduler_latency(tsk, delta >> 10, 0);
1818 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1820 #ifdef CONFIG_SCHED_DEBUG
1821 s64 d = se->vruntime - cfs_rq->min_vruntime;
1826 if (d > 3*sysctl_sched_latency)
1827 schedstat_inc(cfs_rq, nr_spread_over);
1832 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1834 u64 vruntime = cfs_rq->min_vruntime;
1837 * The 'current' period is already promised to the current tasks,
1838 * however the extra weight of the new task will slow them down a
1839 * little, place the new task so that it fits in the slot that
1840 * stays open at the end.
1842 if (initial && sched_feat(START_DEBIT))
1843 vruntime += sched_vslice(cfs_rq, se);
1845 /* sleeps up to a single latency don't count. */
1847 unsigned long thresh = sysctl_sched_latency;
1850 * Halve their sleep time's effect, to allow
1851 * for a gentler effect of sleepers:
1853 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1859 /* ensure we never gain time by being placed backwards. */
1860 se->vruntime = max_vruntime(se->vruntime, vruntime);
1863 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1866 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1869 * Update the normalized vruntime before updating min_vruntime
1870 * through calling update_curr().
1872 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1873 se->vruntime += cfs_rq->min_vruntime;
1876 * Update run-time statistics of the 'current'.
1878 update_curr(cfs_rq);
1879 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1880 account_entity_enqueue(cfs_rq, se);
1881 update_cfs_shares(cfs_rq);
1883 if (flags & ENQUEUE_WAKEUP) {
1884 place_entity(cfs_rq, se, 0);
1885 enqueue_sleeper(cfs_rq, se);
1888 update_stats_enqueue(cfs_rq, se);
1889 check_spread(cfs_rq, se);
1890 if (se != cfs_rq->curr)
1891 __enqueue_entity(cfs_rq, se);
1894 if (cfs_rq->nr_running == 1) {
1895 list_add_leaf_cfs_rq(cfs_rq);
1896 check_enqueue_throttle(cfs_rq);
1900 static void __clear_buddies_last(struct sched_entity *se)
1902 for_each_sched_entity(se) {
1903 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1904 if (cfs_rq->last == se)
1905 cfs_rq->last = NULL;
1911 static void __clear_buddies_next(struct sched_entity *se)
1913 for_each_sched_entity(se) {
1914 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1915 if (cfs_rq->next == se)
1916 cfs_rq->next = NULL;
1922 static void __clear_buddies_skip(struct sched_entity *se)
1924 for_each_sched_entity(se) {
1925 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1926 if (cfs_rq->skip == se)
1927 cfs_rq->skip = NULL;
1933 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1935 if (cfs_rq->last == se)
1936 __clear_buddies_last(se);
1938 if (cfs_rq->next == se)
1939 __clear_buddies_next(se);
1941 if (cfs_rq->skip == se)
1942 __clear_buddies_skip(se);
1945 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1948 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1951 * Update run-time statistics of the 'current'.
1953 update_curr(cfs_rq);
1954 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1956 update_stats_dequeue(cfs_rq, se);
1957 if (flags & DEQUEUE_SLEEP) {
1958 #ifdef CONFIG_SCHEDSTATS
1959 if (entity_is_task(se)) {
1960 struct task_struct *tsk = task_of(se);
1962 if (tsk->state & TASK_INTERRUPTIBLE)
1963 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
1964 if (tsk->state & TASK_UNINTERRUPTIBLE)
1965 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
1970 clear_buddies(cfs_rq, se);
1972 if (se != cfs_rq->curr)
1973 __dequeue_entity(cfs_rq, se);
1975 account_entity_dequeue(cfs_rq, se);
1978 * Normalize the entity after updating the min_vruntime because the
1979 * update can refer to the ->curr item and we need to reflect this
1980 * movement in our normalized position.
1982 if (!(flags & DEQUEUE_SLEEP))
1983 se->vruntime -= cfs_rq->min_vruntime;
1985 /* return excess runtime on last dequeue */
1986 return_cfs_rq_runtime(cfs_rq);
1988 update_min_vruntime(cfs_rq);
1989 update_cfs_shares(cfs_rq);
1993 * Preempt the current task with a newly woken task if needed:
1996 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1998 unsigned long ideal_runtime, delta_exec;
1999 struct sched_entity *se;
2002 ideal_runtime = sched_slice(cfs_rq, curr);
2003 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2004 if (delta_exec > ideal_runtime) {
2005 resched_task(rq_of(cfs_rq)->curr);
2007 * The current task ran long enough, ensure it doesn't get
2008 * re-elected due to buddy favours.
2010 clear_buddies(cfs_rq, curr);
2015 * Ensure that a task that missed wakeup preemption by a
2016 * narrow margin doesn't have to wait for a full slice.
2017 * This also mitigates buddy induced latencies under load.
2019 if (delta_exec < sysctl_sched_min_granularity)
2022 se = __pick_first_entity(cfs_rq);
2023 delta = curr->vruntime - se->vruntime;
2028 if (delta > ideal_runtime)
2029 resched_task(rq_of(cfs_rq)->curr);
2033 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2035 /* 'current' is not kept within the tree. */
2038 * Any task has to be enqueued before it get to execute on
2039 * a CPU. So account for the time it spent waiting on the
2042 update_stats_wait_end(cfs_rq, se);
2043 __dequeue_entity(cfs_rq, se);
2046 update_stats_curr_start(cfs_rq, se);
2048 #ifdef CONFIG_SCHEDSTATS
2050 * Track our maximum slice length, if the CPU's load is at
2051 * least twice that of our own weight (i.e. dont track it
2052 * when there are only lesser-weight tasks around):
2054 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2055 se->statistics.slice_max = max(se->statistics.slice_max,
2056 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2059 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2063 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2066 * Pick the next process, keeping these things in mind, in this order:
2067 * 1) keep things fair between processes/task groups
2068 * 2) pick the "next" process, since someone really wants that to run
2069 * 3) pick the "last" process, for cache locality
2070 * 4) do not run the "skip" process, if something else is available
2072 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2074 struct sched_entity *se = __pick_first_entity(cfs_rq);
2075 struct sched_entity *left = se;
2078 * Avoid running the skip buddy, if running something else can
2079 * be done without getting too unfair.
2081 if (cfs_rq->skip == se) {
2082 struct sched_entity *second = __pick_next_entity(se);
2083 if (second && wakeup_preempt_entity(second, left) < 1)
2088 * Prefer last buddy, try to return the CPU to a preempted task.
2090 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2094 * Someone really wants this to run. If it's not unfair, run it.
2096 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2099 clear_buddies(cfs_rq, se);
2104 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2106 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2109 * If still on the runqueue then deactivate_task()
2110 * was not called and update_curr() has to be done:
2113 update_curr(cfs_rq);
2115 /* throttle cfs_rqs exceeding runtime */
2116 check_cfs_rq_runtime(cfs_rq);
2118 check_spread(cfs_rq, prev);
2120 update_stats_wait_start(cfs_rq, prev);
2121 /* Put 'current' back into the tree. */
2122 __enqueue_entity(cfs_rq, prev);
2123 /* in !on_rq case, update occurred at dequeue */
2124 update_entity_load_avg(prev, 1);
2126 cfs_rq->curr = NULL;
2130 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2133 * Update run-time statistics of the 'current'.
2135 update_curr(cfs_rq);
2138 * Ensure that runnable average is periodically updated.
2140 update_entity_load_avg(curr, 1);
2141 update_cfs_rq_blocked_load(cfs_rq, 1);
2142 update_cfs_shares(cfs_rq);
2144 #ifdef CONFIG_SCHED_HRTICK
2146 * queued ticks are scheduled to match the slice, so don't bother
2147 * validating it and just reschedule.
2150 resched_task(rq_of(cfs_rq)->curr);
2154 * don't let the period tick interfere with the hrtick preemption
2156 if (!sched_feat(DOUBLE_TICK) &&
2157 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2161 if (cfs_rq->nr_running > 1)
2162 check_preempt_tick(cfs_rq, curr);
2166 /**************************************************
2167 * CFS bandwidth control machinery
2170 #ifdef CONFIG_CFS_BANDWIDTH
2172 #ifdef HAVE_JUMP_LABEL
2173 static struct static_key __cfs_bandwidth_used;
2175 static inline bool cfs_bandwidth_used(void)
2177 return static_key_false(&__cfs_bandwidth_used);
2180 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2182 /* only need to count groups transitioning between enabled/!enabled */
2183 if (enabled && !was_enabled)
2184 static_key_slow_inc(&__cfs_bandwidth_used);
2185 else if (!enabled && was_enabled)
2186 static_key_slow_dec(&__cfs_bandwidth_used);
2188 #else /* HAVE_JUMP_LABEL */
2189 static bool cfs_bandwidth_used(void)
2194 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2195 #endif /* HAVE_JUMP_LABEL */
2198 * default period for cfs group bandwidth.
2199 * default: 0.1s, units: nanoseconds
2201 static inline u64 default_cfs_period(void)
2203 return 100000000ULL;
2206 static inline u64 sched_cfs_bandwidth_slice(void)
2208 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2212 * Replenish runtime according to assigned quota and update expiration time.
2213 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2214 * additional synchronization around rq->lock.
2216 * requires cfs_b->lock
2218 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2222 if (cfs_b->quota == RUNTIME_INF)
2225 now = sched_clock_cpu(smp_processor_id());
2226 cfs_b->runtime = cfs_b->quota;
2227 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2230 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2232 return &tg->cfs_bandwidth;
2235 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2236 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2238 if (unlikely(cfs_rq->throttle_count))
2239 return cfs_rq->throttled_clock_task;
2241 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2244 /* returns 0 on failure to allocate runtime */
2245 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2247 struct task_group *tg = cfs_rq->tg;
2248 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2249 u64 amount = 0, min_amount, expires;
2251 /* note: this is a positive sum as runtime_remaining <= 0 */
2252 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2254 raw_spin_lock(&cfs_b->lock);
2255 if (cfs_b->quota == RUNTIME_INF)
2256 amount = min_amount;
2259 * If the bandwidth pool has become inactive, then at least one
2260 * period must have elapsed since the last consumption.
2261 * Refresh the global state and ensure bandwidth timer becomes
2264 if (!cfs_b->timer_active) {
2265 __refill_cfs_bandwidth_runtime(cfs_b);
2266 __start_cfs_bandwidth(cfs_b);
2269 if (cfs_b->runtime > 0) {
2270 amount = min(cfs_b->runtime, min_amount);
2271 cfs_b->runtime -= amount;
2275 expires = cfs_b->runtime_expires;
2276 raw_spin_unlock(&cfs_b->lock);
2278 cfs_rq->runtime_remaining += amount;
2280 * we may have advanced our local expiration to account for allowed
2281 * spread between our sched_clock and the one on which runtime was
2284 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2285 cfs_rq->runtime_expires = expires;
2287 return cfs_rq->runtime_remaining > 0;
2291 * Note: This depends on the synchronization provided by sched_clock and the
2292 * fact that rq->clock snapshots this value.
2294 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2296 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2298 /* if the deadline is ahead of our clock, nothing to do */
2299 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2302 if (cfs_rq->runtime_remaining < 0)
2306 * If the local deadline has passed we have to consider the
2307 * possibility that our sched_clock is 'fast' and the global deadline
2308 * has not truly expired.
2310 * Fortunately we can check determine whether this the case by checking
2311 * whether the global deadline has advanced.
2314 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2315 /* extend local deadline, drift is bounded above by 2 ticks */
2316 cfs_rq->runtime_expires += TICK_NSEC;
2318 /* global deadline is ahead, expiration has passed */
2319 cfs_rq->runtime_remaining = 0;
2323 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2324 unsigned long delta_exec)
2326 /* dock delta_exec before expiring quota (as it could span periods) */
2327 cfs_rq->runtime_remaining -= delta_exec;
2328 expire_cfs_rq_runtime(cfs_rq);
2330 if (likely(cfs_rq->runtime_remaining > 0))
2334 * if we're unable to extend our runtime we resched so that the active
2335 * hierarchy can be throttled
2337 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2338 resched_task(rq_of(cfs_rq)->curr);
2341 static __always_inline
2342 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2344 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2347 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2350 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2352 return cfs_bandwidth_used() && cfs_rq->throttled;
2355 /* check whether cfs_rq, or any parent, is throttled */
2356 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2358 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2362 * Ensure that neither of the group entities corresponding to src_cpu or
2363 * dest_cpu are members of a throttled hierarchy when performing group
2364 * load-balance operations.
2366 static inline int throttled_lb_pair(struct task_group *tg,
2367 int src_cpu, int dest_cpu)
2369 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2371 src_cfs_rq = tg->cfs_rq[src_cpu];
2372 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2374 return throttled_hierarchy(src_cfs_rq) ||
2375 throttled_hierarchy(dest_cfs_rq);
2378 /* updated child weight may affect parent so we have to do this bottom up */
2379 static int tg_unthrottle_up(struct task_group *tg, void *data)
2381 struct rq *rq = data;
2382 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2384 cfs_rq->throttle_count--;
2386 if (!cfs_rq->throttle_count) {
2387 /* adjust cfs_rq_clock_task() */
2388 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2389 cfs_rq->throttled_clock_task;
2396 static int tg_throttle_down(struct task_group *tg, void *data)
2398 struct rq *rq = data;
2399 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2401 /* group is entering throttled state, stop time */
2402 if (!cfs_rq->throttle_count)
2403 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2404 cfs_rq->throttle_count++;
2409 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2411 struct rq *rq = rq_of(cfs_rq);
2412 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2413 struct sched_entity *se;
2414 long task_delta, dequeue = 1;
2416 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2418 /* freeze hierarchy runnable averages while throttled */
2420 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2423 task_delta = cfs_rq->h_nr_running;
2424 for_each_sched_entity(se) {
2425 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2426 /* throttled entity or throttle-on-deactivate */
2431 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2432 qcfs_rq->h_nr_running -= task_delta;
2434 if (qcfs_rq->load.weight)
2439 rq->nr_running -= task_delta;
2441 cfs_rq->throttled = 1;
2442 cfs_rq->throttled_clock = rq_clock(rq);
2443 raw_spin_lock(&cfs_b->lock);
2444 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2445 raw_spin_unlock(&cfs_b->lock);
2448 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2450 struct rq *rq = rq_of(cfs_rq);
2451 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2452 struct sched_entity *se;
2456 se = cfs_rq->tg->se[cpu_of(rq)];
2458 cfs_rq->throttled = 0;
2460 update_rq_clock(rq);
2462 raw_spin_lock(&cfs_b->lock);
2463 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2464 list_del_rcu(&cfs_rq->throttled_list);
2465 raw_spin_unlock(&cfs_b->lock);
2467 /* update hierarchical throttle state */
2468 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2470 if (!cfs_rq->load.weight)
2473 task_delta = cfs_rq->h_nr_running;
2474 for_each_sched_entity(se) {
2478 cfs_rq = cfs_rq_of(se);
2480 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2481 cfs_rq->h_nr_running += task_delta;
2483 if (cfs_rq_throttled(cfs_rq))
2488 rq->nr_running += task_delta;
2490 /* determine whether we need to wake up potentially idle cpu */
2491 if (rq->curr == rq->idle && rq->cfs.nr_running)
2492 resched_task(rq->curr);
2495 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2496 u64 remaining, u64 expires)
2498 struct cfs_rq *cfs_rq;
2499 u64 runtime = remaining;
2502 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2504 struct rq *rq = rq_of(cfs_rq);
2506 raw_spin_lock(&rq->lock);
2507 if (!cfs_rq_throttled(cfs_rq))
2510 runtime = -cfs_rq->runtime_remaining + 1;
2511 if (runtime > remaining)
2512 runtime = remaining;
2513 remaining -= runtime;
2515 cfs_rq->runtime_remaining += runtime;
2516 cfs_rq->runtime_expires = expires;
2518 /* we check whether we're throttled above */
2519 if (cfs_rq->runtime_remaining > 0)
2520 unthrottle_cfs_rq(cfs_rq);
2523 raw_spin_unlock(&rq->lock);
2534 * Responsible for refilling a task_group's bandwidth and unthrottling its
2535 * cfs_rqs as appropriate. If there has been no activity within the last
2536 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2537 * used to track this state.
2539 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2541 u64 runtime, runtime_expires;
2542 int idle = 1, throttled;
2544 raw_spin_lock(&cfs_b->lock);
2545 /* no need to continue the timer with no bandwidth constraint */
2546 if (cfs_b->quota == RUNTIME_INF)
2549 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2550 /* idle depends on !throttled (for the case of a large deficit) */
2551 idle = cfs_b->idle && !throttled;
2552 cfs_b->nr_periods += overrun;
2554 /* if we're going inactive then everything else can be deferred */
2558 __refill_cfs_bandwidth_runtime(cfs_b);
2561 /* mark as potentially idle for the upcoming period */
2566 /* account preceding periods in which throttling occurred */
2567 cfs_b->nr_throttled += overrun;
2570 * There are throttled entities so we must first use the new bandwidth
2571 * to unthrottle them before making it generally available. This
2572 * ensures that all existing debts will be paid before a new cfs_rq is
2575 runtime = cfs_b->runtime;
2576 runtime_expires = cfs_b->runtime_expires;
2580 * This check is repeated as we are holding onto the new bandwidth
2581 * while we unthrottle. This can potentially race with an unthrottled
2582 * group trying to acquire new bandwidth from the global pool.
2584 while (throttled && runtime > 0) {
2585 raw_spin_unlock(&cfs_b->lock);
2586 /* we can't nest cfs_b->lock while distributing bandwidth */
2587 runtime = distribute_cfs_runtime(cfs_b, runtime,
2589 raw_spin_lock(&cfs_b->lock);
2591 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2594 /* return (any) remaining runtime */
2595 cfs_b->runtime = runtime;
2597 * While we are ensured activity in the period following an
2598 * unthrottle, this also covers the case in which the new bandwidth is
2599 * insufficient to cover the existing bandwidth deficit. (Forcing the
2600 * timer to remain active while there are any throttled entities.)
2605 cfs_b->timer_active = 0;
2606 raw_spin_unlock(&cfs_b->lock);
2611 /* a cfs_rq won't donate quota below this amount */
2612 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2613 /* minimum remaining period time to redistribute slack quota */
2614 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2615 /* how long we wait to gather additional slack before distributing */
2616 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2618 /* are we near the end of the current quota period? */
2619 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2621 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2624 /* if the call-back is running a quota refresh is already occurring */
2625 if (hrtimer_callback_running(refresh_timer))
2628 /* is a quota refresh about to occur? */
2629 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2630 if (remaining < min_expire)
2636 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2638 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2640 /* if there's a quota refresh soon don't bother with slack */
2641 if (runtime_refresh_within(cfs_b, min_left))
2644 start_bandwidth_timer(&cfs_b->slack_timer,
2645 ns_to_ktime(cfs_bandwidth_slack_period));
2648 /* we know any runtime found here is valid as update_curr() precedes return */
2649 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2651 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2652 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2654 if (slack_runtime <= 0)
2657 raw_spin_lock(&cfs_b->lock);
2658 if (cfs_b->quota != RUNTIME_INF &&
2659 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2660 cfs_b->runtime += slack_runtime;
2662 /* we are under rq->lock, defer unthrottling using a timer */
2663 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2664 !list_empty(&cfs_b->throttled_cfs_rq))
2665 start_cfs_slack_bandwidth(cfs_b);
2667 raw_spin_unlock(&cfs_b->lock);
2669 /* even if it's not valid for return we don't want to try again */
2670 cfs_rq->runtime_remaining -= slack_runtime;
2673 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2675 if (!cfs_bandwidth_used())
2678 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2681 __return_cfs_rq_runtime(cfs_rq);
2685 * This is done with a timer (instead of inline with bandwidth return) since
2686 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2688 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2690 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2693 /* confirm we're still not at a refresh boundary */
2694 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2697 raw_spin_lock(&cfs_b->lock);
2698 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2699 runtime = cfs_b->runtime;
2702 expires = cfs_b->runtime_expires;
2703 raw_spin_unlock(&cfs_b->lock);
2708 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2710 raw_spin_lock(&cfs_b->lock);
2711 if (expires == cfs_b->runtime_expires)
2712 cfs_b->runtime = runtime;
2713 raw_spin_unlock(&cfs_b->lock);
2717 * When a group wakes up we want to make sure that its quota is not already
2718 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2719 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2721 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2723 if (!cfs_bandwidth_used())
2726 /* an active group must be handled by the update_curr()->put() path */
2727 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2730 /* ensure the group is not already throttled */
2731 if (cfs_rq_throttled(cfs_rq))
2734 /* update runtime allocation */
2735 account_cfs_rq_runtime(cfs_rq, 0);
2736 if (cfs_rq->runtime_remaining <= 0)
2737 throttle_cfs_rq(cfs_rq);
2740 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2741 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2743 if (!cfs_bandwidth_used())
2746 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2750 * it's possible for a throttled entity to be forced into a running
2751 * state (e.g. set_curr_task), in this case we're finished.
2753 if (cfs_rq_throttled(cfs_rq))
2756 throttle_cfs_rq(cfs_rq);
2759 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2761 struct cfs_bandwidth *cfs_b =
2762 container_of(timer, struct cfs_bandwidth, slack_timer);
2763 do_sched_cfs_slack_timer(cfs_b);
2765 return HRTIMER_NORESTART;
2768 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2770 struct cfs_bandwidth *cfs_b =
2771 container_of(timer, struct cfs_bandwidth, period_timer);
2777 now = hrtimer_cb_get_time(timer);
2778 overrun = hrtimer_forward(timer, now, cfs_b->period);
2783 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2786 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2789 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2791 raw_spin_lock_init(&cfs_b->lock);
2793 cfs_b->quota = RUNTIME_INF;
2794 cfs_b->period = ns_to_ktime(default_cfs_period());
2796 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2797 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2798 cfs_b->period_timer.function = sched_cfs_period_timer;
2799 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2800 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2803 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2805 cfs_rq->runtime_enabled = 0;
2806 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2809 /* requires cfs_b->lock, may release to reprogram timer */
2810 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2813 * The timer may be active because we're trying to set a new bandwidth
2814 * period or because we're racing with the tear-down path
2815 * (timer_active==0 becomes visible before the hrtimer call-back
2816 * terminates). In either case we ensure that it's re-programmed
2818 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2819 raw_spin_unlock(&cfs_b->lock);
2820 /* ensure cfs_b->lock is available while we wait */
2821 hrtimer_cancel(&cfs_b->period_timer);
2823 raw_spin_lock(&cfs_b->lock);
2824 /* if someone else restarted the timer then we're done */
2825 if (cfs_b->timer_active)
2829 cfs_b->timer_active = 1;
2830 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2833 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2835 hrtimer_cancel(&cfs_b->period_timer);
2836 hrtimer_cancel(&cfs_b->slack_timer);
2839 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2841 struct cfs_rq *cfs_rq;
2843 for_each_leaf_cfs_rq(rq, cfs_rq) {
2844 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2846 if (!cfs_rq->runtime_enabled)
2850 * clock_task is not advancing so we just need to make sure
2851 * there's some valid quota amount
2853 cfs_rq->runtime_remaining = cfs_b->quota;
2854 if (cfs_rq_throttled(cfs_rq))
2855 unthrottle_cfs_rq(cfs_rq);
2859 #else /* CONFIG_CFS_BANDWIDTH */
2860 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2862 return rq_clock_task(rq_of(cfs_rq));
2865 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2866 unsigned long delta_exec) {}
2867 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2868 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2869 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2871 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2876 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2881 static inline int throttled_lb_pair(struct task_group *tg,
2882 int src_cpu, int dest_cpu)
2887 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2889 #ifdef CONFIG_FAIR_GROUP_SCHED
2890 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2893 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2897 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2898 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2900 #endif /* CONFIG_CFS_BANDWIDTH */
2902 /**************************************************
2903 * CFS operations on tasks:
2906 #ifdef CONFIG_SCHED_HRTICK
2907 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2909 struct sched_entity *se = &p->se;
2910 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912 WARN_ON(task_rq(p) != rq);
2914 if (cfs_rq->nr_running > 1) {
2915 u64 slice = sched_slice(cfs_rq, se);
2916 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2917 s64 delta = slice - ran;
2926 * Don't schedule slices shorter than 10000ns, that just
2927 * doesn't make sense. Rely on vruntime for fairness.
2930 delta = max_t(s64, 10000LL, delta);
2932 hrtick_start(rq, delta);
2937 * called from enqueue/dequeue and updates the hrtick when the
2938 * current task is from our class and nr_running is low enough
2941 static void hrtick_update(struct rq *rq)
2943 struct task_struct *curr = rq->curr;
2945 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2948 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2949 hrtick_start_fair(rq, curr);
2951 #else /* !CONFIG_SCHED_HRTICK */
2953 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2957 static inline void hrtick_update(struct rq *rq)
2963 * The enqueue_task method is called before nr_running is
2964 * increased. Here we update the fair scheduling stats and
2965 * then put the task into the rbtree:
2968 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2970 struct cfs_rq *cfs_rq;
2971 struct sched_entity *se = &p->se;
2973 for_each_sched_entity(se) {
2976 cfs_rq = cfs_rq_of(se);
2977 enqueue_entity(cfs_rq, se, flags);
2980 * end evaluation on encountering a throttled cfs_rq
2982 * note: in the case of encountering a throttled cfs_rq we will
2983 * post the final h_nr_running increment below.
2985 if (cfs_rq_throttled(cfs_rq))
2987 cfs_rq->h_nr_running++;
2989 flags = ENQUEUE_WAKEUP;
2992 for_each_sched_entity(se) {
2993 cfs_rq = cfs_rq_of(se);
2994 cfs_rq->h_nr_running++;
2996 if (cfs_rq_throttled(cfs_rq))
2999 update_cfs_shares(cfs_rq);
3000 update_entity_load_avg(se, 1);
3004 update_rq_runnable_avg(rq, rq->nr_running);
3010 static void set_next_buddy(struct sched_entity *se);
3013 * The dequeue_task method is called before nr_running is
3014 * decreased. We remove the task from the rbtree and
3015 * update the fair scheduling stats:
3017 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3019 struct cfs_rq *cfs_rq;
3020 struct sched_entity *se = &p->se;
3021 int task_sleep = flags & DEQUEUE_SLEEP;
3023 for_each_sched_entity(se) {
3024 cfs_rq = cfs_rq_of(se);
3025 dequeue_entity(cfs_rq, se, flags);
3028 * end evaluation on encountering a throttled cfs_rq
3030 * note: in the case of encountering a throttled cfs_rq we will
3031 * post the final h_nr_running decrement below.
3033 if (cfs_rq_throttled(cfs_rq))
3035 cfs_rq->h_nr_running--;
3037 /* Don't dequeue parent if it has other entities besides us */
3038 if (cfs_rq->load.weight) {
3040 * Bias pick_next to pick a task from this cfs_rq, as
3041 * p is sleeping when it is within its sched_slice.
3043 if (task_sleep && parent_entity(se))
3044 set_next_buddy(parent_entity(se));
3046 /* avoid re-evaluating load for this entity */
3047 se = parent_entity(se);
3050 flags |= DEQUEUE_SLEEP;
3053 for_each_sched_entity(se) {
3054 cfs_rq = cfs_rq_of(se);
3055 cfs_rq->h_nr_running--;
3057 if (cfs_rq_throttled(cfs_rq))
3060 update_cfs_shares(cfs_rq);
3061 update_entity_load_avg(se, 1);
3066 update_rq_runnable_avg(rq, 1);
3072 /* Used instead of source_load when we know the type == 0 */
3073 static unsigned long weighted_cpuload(const int cpu)
3075 return cpu_rq(cpu)->cfs.runnable_load_avg;
3079 * Return a low guess at the load of a migration-source cpu weighted
3080 * according to the scheduling class and "nice" value.
3082 * We want to under-estimate the load of migration sources, to
3083 * balance conservatively.
3085 static unsigned long source_load(int cpu, int type)
3087 struct rq *rq = cpu_rq(cpu);
3088 unsigned long total = weighted_cpuload(cpu);
3090 if (type == 0 || !sched_feat(LB_BIAS))
3093 return min(rq->cpu_load[type-1], total);
3097 * Return a high guess at the load of a migration-target cpu weighted
3098 * according to the scheduling class and "nice" value.
3100 static unsigned long target_load(int cpu, int type)
3102 struct rq *rq = cpu_rq(cpu);
3103 unsigned long total = weighted_cpuload(cpu);
3105 if (type == 0 || !sched_feat(LB_BIAS))
3108 return max(rq->cpu_load[type-1], total);
3111 static unsigned long power_of(int cpu)
3113 return cpu_rq(cpu)->cpu_power;
3116 static unsigned long cpu_avg_load_per_task(int cpu)
3118 struct rq *rq = cpu_rq(cpu);
3119 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3120 unsigned long load_avg = rq->cfs.runnable_load_avg;
3123 return load_avg / nr_running;
3128 static void record_wakee(struct task_struct *p)
3131 * Rough decay (wiping) for cost saving, don't worry
3132 * about the boundary, really active task won't care
3135 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3136 current->wakee_flips = 0;
3137 current->wakee_flip_decay_ts = jiffies;
3140 if (current->last_wakee != p) {
3141 current->last_wakee = p;
3142 current->wakee_flips++;
3146 static void task_waking_fair(struct task_struct *p)
3148 struct sched_entity *se = &p->se;
3149 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3152 #ifndef CONFIG_64BIT
3153 u64 min_vruntime_copy;
3156 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3158 min_vruntime = cfs_rq->min_vruntime;
3159 } while (min_vruntime != min_vruntime_copy);
3161 min_vruntime = cfs_rq->min_vruntime;
3164 se->vruntime -= min_vruntime;
3168 #ifdef CONFIG_FAIR_GROUP_SCHED
3170 * effective_load() calculates the load change as seen from the root_task_group
3172 * Adding load to a group doesn't make a group heavier, but can cause movement
3173 * of group shares between cpus. Assuming the shares were perfectly aligned one
3174 * can calculate the shift in shares.
3176 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3177 * on this @cpu and results in a total addition (subtraction) of @wg to the
3178 * total group weight.
3180 * Given a runqueue weight distribution (rw_i) we can compute a shares
3181 * distribution (s_i) using:
3183 * s_i = rw_i / \Sum rw_j (1)
3185 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3186 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3187 * shares distribution (s_i):
3189 * rw_i = { 2, 4, 1, 0 }
3190 * s_i = { 2/7, 4/7, 1/7, 0 }
3192 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3193 * task used to run on and the CPU the waker is running on), we need to
3194 * compute the effect of waking a task on either CPU and, in case of a sync
3195 * wakeup, compute the effect of the current task going to sleep.
3197 * So for a change of @wl to the local @cpu with an overall group weight change
3198 * of @wl we can compute the new shares distribution (s'_i) using:
3200 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3202 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3203 * differences in waking a task to CPU 0. The additional task changes the
3204 * weight and shares distributions like:
3206 * rw'_i = { 3, 4, 1, 0 }
3207 * s'_i = { 3/8, 4/8, 1/8, 0 }
3209 * We can then compute the difference in effective weight by using:
3211 * dw_i = S * (s'_i - s_i) (3)
3213 * Where 'S' is the group weight as seen by its parent.
3215 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3216 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3217 * 4/7) times the weight of the group.
3219 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3221 struct sched_entity *se = tg->se[cpu];
3223 if (!tg->parent) /* the trivial, non-cgroup case */
3226 for_each_sched_entity(se) {
3232 * W = @wg + \Sum rw_j
3234 W = wg + calc_tg_weight(tg, se->my_q);
3239 w = se->my_q->load.weight + wl;
3242 * wl = S * s'_i; see (2)
3245 wl = (w * tg->shares) / W;
3250 * Per the above, wl is the new se->load.weight value; since
3251 * those are clipped to [MIN_SHARES, ...) do so now. See
3252 * calc_cfs_shares().
3254 if (wl < MIN_SHARES)
3258 * wl = dw_i = S * (s'_i - s_i); see (3)
3260 wl -= se->load.weight;
3263 * Recursively apply this logic to all parent groups to compute
3264 * the final effective load change on the root group. Since
3265 * only the @tg group gets extra weight, all parent groups can
3266 * only redistribute existing shares. @wl is the shift in shares
3267 * resulting from this level per the above.
3276 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3277 unsigned long wl, unsigned long wg)
3284 static int wake_wide(struct task_struct *p)
3286 int factor = this_cpu_read(sd_llc_size);
3289 * Yeah, it's the switching-frequency, could means many wakee or
3290 * rapidly switch, use factor here will just help to automatically
3291 * adjust the loose-degree, so bigger node will lead to more pull.
3293 if (p->wakee_flips > factor) {
3295 * wakee is somewhat hot, it needs certain amount of cpu
3296 * resource, so if waker is far more hot, prefer to leave
3299 if (current->wakee_flips > (factor * p->wakee_flips))
3306 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3308 s64 this_load, load;
3309 int idx, this_cpu, prev_cpu;
3310 unsigned long tl_per_task;
3311 struct task_group *tg;
3312 unsigned long weight;
3316 * If we wake multiple tasks be careful to not bounce
3317 * ourselves around too much.
3323 this_cpu = smp_processor_id();
3324 prev_cpu = task_cpu(p);
3325 load = source_load(prev_cpu, idx);
3326 this_load = target_load(this_cpu, idx);
3329 * If sync wakeup then subtract the (maximum possible)
3330 * effect of the currently running task from the load
3331 * of the current CPU:
3334 tg = task_group(current);
3335 weight = current->se.load.weight;
3337 this_load += effective_load(tg, this_cpu, -weight, -weight);
3338 load += effective_load(tg, prev_cpu, 0, -weight);
3342 weight = p->se.load.weight;
3345 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3346 * due to the sync cause above having dropped this_load to 0, we'll
3347 * always have an imbalance, but there's really nothing you can do
3348 * about that, so that's good too.
3350 * Otherwise check if either cpus are near enough in load to allow this
3351 * task to be woken on this_cpu.
3353 if (this_load > 0) {
3354 s64 this_eff_load, prev_eff_load;
3356 this_eff_load = 100;
3357 this_eff_load *= power_of(prev_cpu);
3358 this_eff_load *= this_load +
3359 effective_load(tg, this_cpu, weight, weight);
3361 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3362 prev_eff_load *= power_of(this_cpu);
3363 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3365 balanced = this_eff_load <= prev_eff_load;
3370 * If the currently running task will sleep within
3371 * a reasonable amount of time then attract this newly
3374 if (sync && balanced)
3377 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3378 tl_per_task = cpu_avg_load_per_task(this_cpu);
3381 (this_load <= load &&
3382 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3384 * This domain has SD_WAKE_AFFINE and
3385 * p is cache cold in this domain, and
3386 * there is no bad imbalance.
3388 schedstat_inc(sd, ttwu_move_affine);
3389 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3397 * find_idlest_group finds and returns the least busy CPU group within the
3400 static struct sched_group *
3401 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3402 int this_cpu, int load_idx)
3404 struct sched_group *idlest = NULL, *group = sd->groups;
3405 unsigned long min_load = ULONG_MAX, this_load = 0;
3406 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3409 unsigned long load, avg_load;
3413 /* Skip over this group if it has no CPUs allowed */
3414 if (!cpumask_intersects(sched_group_cpus(group),
3415 tsk_cpus_allowed(p)))
3418 local_group = cpumask_test_cpu(this_cpu,
3419 sched_group_cpus(group));
3421 /* Tally up the load of all CPUs in the group */
3424 for_each_cpu(i, sched_group_cpus(group)) {
3425 /* Bias balancing toward cpus of our domain */
3427 load = source_load(i, load_idx);
3429 load = target_load(i, load_idx);
3434 /* Adjust by relative CPU power of the group */
3435 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3438 this_load = avg_load;
3439 } else if (avg_load < min_load) {
3440 min_load = avg_load;
3443 } while (group = group->next, group != sd->groups);
3445 if (!idlest || 100*this_load < imbalance*min_load)
3451 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3454 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3456 unsigned long load, min_load = ULONG_MAX;
3460 /* Traverse only the allowed CPUs */
3461 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3462 load = weighted_cpuload(i);
3464 if (load < min_load || (load == min_load && i == this_cpu)) {
3474 * Try and locate an idle CPU in the sched_domain.
3476 static int select_idle_sibling(struct task_struct *p, int target)
3478 struct sched_domain *sd;
3479 struct sched_group *sg;
3480 int i = task_cpu(p);
3482 if (idle_cpu(target))
3486 * If the prevous cpu is cache affine and idle, don't be stupid.
3488 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3492 * Otherwise, iterate the domains and find an elegible idle cpu.
3494 sd = rcu_dereference(per_cpu(sd_llc, target));
3495 for_each_lower_domain(sd) {
3498 if (!cpumask_intersects(sched_group_cpus(sg),
3499 tsk_cpus_allowed(p)))
3502 for_each_cpu(i, sched_group_cpus(sg)) {
3503 if (i == target || !idle_cpu(i))
3507 target = cpumask_first_and(sched_group_cpus(sg),
3508 tsk_cpus_allowed(p));
3512 } while (sg != sd->groups);
3519 * sched_balance_self: balance the current task (running on cpu) in domains
3520 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3523 * Balance, ie. select the least loaded group.
3525 * Returns the target CPU number, or the same CPU if no balancing is needed.
3527 * preempt must be disabled.
3530 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3532 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3533 int cpu = smp_processor_id();
3534 int prev_cpu = task_cpu(p);
3536 int want_affine = 0;
3537 int sync = wake_flags & WF_SYNC;
3539 if (p->nr_cpus_allowed == 1)
3542 if (sd_flag & SD_BALANCE_WAKE) {
3543 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3549 for_each_domain(cpu, tmp) {
3550 if (!(tmp->flags & SD_LOAD_BALANCE))
3554 * If both cpu and prev_cpu are part of this domain,
3555 * cpu is a valid SD_WAKE_AFFINE target.
3557 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3558 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3563 if (tmp->flags & sd_flag)
3568 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3571 new_cpu = select_idle_sibling(p, prev_cpu);
3576 int load_idx = sd->forkexec_idx;
3577 struct sched_group *group;
3580 if (!(sd->flags & sd_flag)) {
3585 if (sd_flag & SD_BALANCE_WAKE)
3586 load_idx = sd->wake_idx;
3588 group = find_idlest_group(sd, p, cpu, load_idx);
3594 new_cpu = find_idlest_cpu(group, p, cpu);
3595 if (new_cpu == -1 || new_cpu == cpu) {
3596 /* Now try balancing at a lower domain level of cpu */
3601 /* Now try balancing at a lower domain level of new_cpu */
3603 weight = sd->span_weight;
3605 for_each_domain(cpu, tmp) {
3606 if (weight <= tmp->span_weight)
3608 if (tmp->flags & sd_flag)
3611 /* while loop will break here if sd == NULL */
3620 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3621 * cfs_rq_of(p) references at time of call are still valid and identify the
3622 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3623 * other assumptions, including the state of rq->lock, should be made.
3626 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3628 struct sched_entity *se = &p->se;
3629 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3632 * Load tracking: accumulate removed load so that it can be processed
3633 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3634 * to blocked load iff they have a positive decay-count. It can never
3635 * be negative here since on-rq tasks have decay-count == 0.
3637 if (se->avg.decay_count) {
3638 se->avg.decay_count = -__synchronize_entity_decay(se);
3639 atomic_long_add(se->avg.load_avg_contrib,
3640 &cfs_rq->removed_load);
3643 #endif /* CONFIG_SMP */
3645 static unsigned long
3646 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3648 unsigned long gran = sysctl_sched_wakeup_granularity;
3651 * Since its curr running now, convert the gran from real-time
3652 * to virtual-time in his units.
3654 * By using 'se' instead of 'curr' we penalize light tasks, so
3655 * they get preempted easier. That is, if 'se' < 'curr' then
3656 * the resulting gran will be larger, therefore penalizing the
3657 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3658 * be smaller, again penalizing the lighter task.
3660 * This is especially important for buddies when the leftmost
3661 * task is higher priority than the buddy.
3663 return calc_delta_fair(gran, se);
3667 * Should 'se' preempt 'curr'.
3681 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3683 s64 gran, vdiff = curr->vruntime - se->vruntime;
3688 gran = wakeup_gran(curr, se);
3695 static void set_last_buddy(struct sched_entity *se)
3697 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3700 for_each_sched_entity(se)
3701 cfs_rq_of(se)->last = se;
3704 static void set_next_buddy(struct sched_entity *se)
3706 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3709 for_each_sched_entity(se)
3710 cfs_rq_of(se)->next = se;
3713 static void set_skip_buddy(struct sched_entity *se)
3715 for_each_sched_entity(se)
3716 cfs_rq_of(se)->skip = se;
3720 * Preempt the current task with a newly woken task if needed:
3722 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3724 struct task_struct *curr = rq->curr;
3725 struct sched_entity *se = &curr->se, *pse = &p->se;
3726 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3727 int scale = cfs_rq->nr_running >= sched_nr_latency;
3728 int next_buddy_marked = 0;
3730 if (unlikely(se == pse))
3734 * This is possible from callers such as move_task(), in which we
3735 * unconditionally check_prempt_curr() after an enqueue (which may have
3736 * lead to a throttle). This both saves work and prevents false
3737 * next-buddy nomination below.
3739 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3742 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3743 set_next_buddy(pse);
3744 next_buddy_marked = 1;
3748 * We can come here with TIF_NEED_RESCHED already set from new task
3751 * Note: this also catches the edge-case of curr being in a throttled
3752 * group (e.g. via set_curr_task), since update_curr() (in the
3753 * enqueue of curr) will have resulted in resched being set. This
3754 * prevents us from potentially nominating it as a false LAST_BUDDY
3757 if (test_tsk_need_resched(curr))
3760 /* Idle tasks are by definition preempted by non-idle tasks. */
3761 if (unlikely(curr->policy == SCHED_IDLE) &&
3762 likely(p->policy != SCHED_IDLE))
3766 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3767 * is driven by the tick):
3769 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3772 find_matching_se(&se, &pse);
3773 update_curr(cfs_rq_of(se));
3775 if (wakeup_preempt_entity(se, pse) == 1) {
3777 * Bias pick_next to pick the sched entity that is
3778 * triggering this preemption.
3780 if (!next_buddy_marked)
3781 set_next_buddy(pse);
3790 * Only set the backward buddy when the current task is still
3791 * on the rq. This can happen when a wakeup gets interleaved
3792 * with schedule on the ->pre_schedule() or idle_balance()
3793 * point, either of which can * drop the rq lock.
3795 * Also, during early boot the idle thread is in the fair class,
3796 * for obvious reasons its a bad idea to schedule back to it.
3798 if (unlikely(!se->on_rq || curr == rq->idle))
3801 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3805 static struct task_struct *pick_next_task_fair(struct rq *rq)
3807 struct task_struct *p;
3808 struct cfs_rq *cfs_rq = &rq->cfs;
3809 struct sched_entity *se;
3811 if (!cfs_rq->nr_running)
3815 se = pick_next_entity(cfs_rq);
3816 set_next_entity(cfs_rq, se);
3817 cfs_rq = group_cfs_rq(se);
3821 if (hrtick_enabled(rq))
3822 hrtick_start_fair(rq, p);
3828 * Account for a descheduled task:
3830 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3832 struct sched_entity *se = &prev->se;
3833 struct cfs_rq *cfs_rq;
3835 for_each_sched_entity(se) {
3836 cfs_rq = cfs_rq_of(se);
3837 put_prev_entity(cfs_rq, se);
3842 * sched_yield() is very simple
3844 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3846 static void yield_task_fair(struct rq *rq)
3848 struct task_struct *curr = rq->curr;
3849 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3850 struct sched_entity *se = &curr->se;
3853 * Are we the only task in the tree?
3855 if (unlikely(rq->nr_running == 1))
3858 clear_buddies(cfs_rq, se);
3860 if (curr->policy != SCHED_BATCH) {
3861 update_rq_clock(rq);
3863 * Update run-time statistics of the 'current'.
3865 update_curr(cfs_rq);
3867 * Tell update_rq_clock() that we've just updated,
3868 * so we don't do microscopic update in schedule()
3869 * and double the fastpath cost.
3871 rq->skip_clock_update = 1;
3877 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3879 struct sched_entity *se = &p->se;
3881 /* throttled hierarchies are not runnable */
3882 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3885 /* Tell the scheduler that we'd really like pse to run next. */
3888 yield_task_fair(rq);
3894 /**************************************************
3895 * Fair scheduling class load-balancing methods.
3899 * The purpose of load-balancing is to achieve the same basic fairness the
3900 * per-cpu scheduler provides, namely provide a proportional amount of compute
3901 * time to each task. This is expressed in the following equation:
3903 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3905 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3906 * W_i,0 is defined as:
3908 * W_i,0 = \Sum_j w_i,j (2)
3910 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3911 * is derived from the nice value as per prio_to_weight[].
3913 * The weight average is an exponential decay average of the instantaneous
3916 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3918 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3919 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3920 * can also include other factors [XXX].
3922 * To achieve this balance we define a measure of imbalance which follows
3923 * directly from (1):
3925 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3927 * We them move tasks around to minimize the imbalance. In the continuous
3928 * function space it is obvious this converges, in the discrete case we get
3929 * a few fun cases generally called infeasible weight scenarios.
3932 * - infeasible weights;
3933 * - local vs global optima in the discrete case. ]
3938 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3939 * for all i,j solution, we create a tree of cpus that follows the hardware
3940 * topology where each level pairs two lower groups (or better). This results
3941 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3942 * tree to only the first of the previous level and we decrease the frequency
3943 * of load-balance at each level inv. proportional to the number of cpus in
3949 * \Sum { --- * --- * 2^i } = O(n) (5)
3951 * `- size of each group
3952 * | | `- number of cpus doing load-balance
3954 * `- sum over all levels
3956 * Coupled with a limit on how many tasks we can migrate every balance pass,
3957 * this makes (5) the runtime complexity of the balancer.
3959 * An important property here is that each CPU is still (indirectly) connected
3960 * to every other cpu in at most O(log n) steps:
3962 * The adjacency matrix of the resulting graph is given by:
3965 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3968 * And you'll find that:
3970 * A^(log_2 n)_i,j != 0 for all i,j (7)
3972 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3973 * The task movement gives a factor of O(m), giving a convergence complexity
3976 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3981 * In order to avoid CPUs going idle while there's still work to do, new idle
3982 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3983 * tree itself instead of relying on other CPUs to bring it work.
3985 * This adds some complexity to both (5) and (8) but it reduces the total idle
3993 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3996 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4001 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4003 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4005 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4008 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4009 * rewrite all of this once again.]
4012 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4014 #define LBF_ALL_PINNED 0x01
4015 #define LBF_NEED_BREAK 0x02
4016 #define LBF_DST_PINNED 0x04
4017 #define LBF_SOME_PINNED 0x08
4020 struct sched_domain *sd;
4028 struct cpumask *dst_grpmask;
4030 enum cpu_idle_type idle;
4032 /* The set of CPUs under consideration for load-balancing */
4033 struct cpumask *cpus;
4038 unsigned int loop_break;
4039 unsigned int loop_max;
4043 * move_task - move a task from one runqueue to another runqueue.
4044 * Both runqueues must be locked.
4046 static void move_task(struct task_struct *p, struct lb_env *env)
4048 deactivate_task(env->src_rq, p, 0);
4049 set_task_cpu(p, env->dst_cpu);
4050 activate_task(env->dst_rq, p, 0);
4051 check_preempt_curr(env->dst_rq, p, 0);
4055 * Is this task likely cache-hot:
4058 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4062 if (p->sched_class != &fair_sched_class)
4065 if (unlikely(p->policy == SCHED_IDLE))
4069 * Buddy candidates are cache hot:
4071 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4072 (&p->se == cfs_rq_of(&p->se)->next ||
4073 &p->se == cfs_rq_of(&p->se)->last))
4076 if (sysctl_sched_migration_cost == -1)
4078 if (sysctl_sched_migration_cost == 0)
4081 delta = now - p->se.exec_start;
4083 return delta < (s64)sysctl_sched_migration_cost;
4086 #ifdef CONFIG_NUMA_BALANCING
4087 /* Returns true if the destination node has incurred more faults */
4088 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4090 int src_nid, dst_nid;
4092 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4093 !(env->sd->flags & SD_NUMA)) {
4097 src_nid = cpu_to_node(env->src_cpu);
4098 dst_nid = cpu_to_node(env->dst_cpu);
4100 if (src_nid == dst_nid ||
4101 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4104 if (dst_nid == p->numa_preferred_nid ||
4105 p->numa_faults[dst_nid] > p->numa_faults[src_nid])
4112 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4114 int src_nid, dst_nid;
4116 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4119 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4122 src_nid = cpu_to_node(env->src_cpu);
4123 dst_nid = cpu_to_node(env->dst_cpu);
4125 if (src_nid == dst_nid ||
4126 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4129 if (p->numa_faults[dst_nid] < p->numa_faults[src_nid])
4136 static inline bool migrate_improves_locality(struct task_struct *p,
4142 static inline bool migrate_degrades_locality(struct task_struct *p,
4150 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4153 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4155 int tsk_cache_hot = 0;
4157 * We do not migrate tasks that are:
4158 * 1) throttled_lb_pair, or
4159 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4160 * 3) running (obviously), or
4161 * 4) are cache-hot on their current CPU.
4163 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4166 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4169 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4171 env->flags |= LBF_SOME_PINNED;
4174 * Remember if this task can be migrated to any other cpu in
4175 * our sched_group. We may want to revisit it if we couldn't
4176 * meet load balance goals by pulling other tasks on src_cpu.
4178 * Also avoid computing new_dst_cpu if we have already computed
4179 * one in current iteration.
4181 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4184 /* Prevent to re-select dst_cpu via env's cpus */
4185 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4186 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4187 env->flags |= LBF_DST_PINNED;
4188 env->new_dst_cpu = cpu;
4196 /* Record that we found atleast one task that could run on dst_cpu */
4197 env->flags &= ~LBF_ALL_PINNED;
4199 if (task_running(env->src_rq, p)) {
4200 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4205 * Aggressive migration if:
4206 * 1) destination numa is preferred
4207 * 2) task is cache cold, or
4208 * 3) too many balance attempts have failed.
4210 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4212 tsk_cache_hot = migrate_degrades_locality(p, env);
4214 if (migrate_improves_locality(p, env)) {
4215 #ifdef CONFIG_SCHEDSTATS
4216 if (tsk_cache_hot) {
4217 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4218 schedstat_inc(p, se.statistics.nr_forced_migrations);
4224 if (!tsk_cache_hot ||
4225 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4227 if (tsk_cache_hot) {
4228 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4229 schedstat_inc(p, se.statistics.nr_forced_migrations);
4235 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4240 * move_one_task tries to move exactly one task from busiest to this_rq, as
4241 * part of active balancing operations within "domain".
4242 * Returns 1 if successful and 0 otherwise.
4244 * Called with both runqueues locked.
4246 static int move_one_task(struct lb_env *env)
4248 struct task_struct *p, *n;
4250 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4251 if (!can_migrate_task(p, env))
4256 * Right now, this is only the second place move_task()
4257 * is called, so we can safely collect move_task()
4258 * stats here rather than inside move_task().
4260 schedstat_inc(env->sd, lb_gained[env->idle]);
4266 static unsigned long task_h_load(struct task_struct *p);
4268 static const unsigned int sched_nr_migrate_break = 32;
4271 * move_tasks tries to move up to imbalance weighted load from busiest to
4272 * this_rq, as part of a balancing operation within domain "sd".
4273 * Returns 1 if successful and 0 otherwise.
4275 * Called with both runqueues locked.
4277 static int move_tasks(struct lb_env *env)
4279 struct list_head *tasks = &env->src_rq->cfs_tasks;
4280 struct task_struct *p;
4284 if (env->imbalance <= 0)
4287 while (!list_empty(tasks)) {
4288 p = list_first_entry(tasks, struct task_struct, se.group_node);
4291 /* We've more or less seen every task there is, call it quits */
4292 if (env->loop > env->loop_max)
4295 /* take a breather every nr_migrate tasks */
4296 if (env->loop > env->loop_break) {
4297 env->loop_break += sched_nr_migrate_break;
4298 env->flags |= LBF_NEED_BREAK;
4302 if (!can_migrate_task(p, env))
4305 load = task_h_load(p);
4307 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4310 if ((load / 2) > env->imbalance)
4315 env->imbalance -= load;
4317 #ifdef CONFIG_PREEMPT
4319 * NEWIDLE balancing is a source of latency, so preemptible
4320 * kernels will stop after the first task is pulled to minimize
4321 * the critical section.
4323 if (env->idle == CPU_NEWLY_IDLE)
4328 * We only want to steal up to the prescribed amount of
4331 if (env->imbalance <= 0)
4336 list_move_tail(&p->se.group_node, tasks);
4340 * Right now, this is one of only two places move_task() is called,
4341 * so we can safely collect move_task() stats here rather than
4342 * inside move_task().
4344 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4349 #ifdef CONFIG_FAIR_GROUP_SCHED
4351 * update tg->load_weight by folding this cpu's load_avg
4353 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4355 struct sched_entity *se = tg->se[cpu];
4356 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4358 /* throttled entities do not contribute to load */
4359 if (throttled_hierarchy(cfs_rq))
4362 update_cfs_rq_blocked_load(cfs_rq, 1);
4365 update_entity_load_avg(se, 1);
4367 * We pivot on our runnable average having decayed to zero for
4368 * list removal. This generally implies that all our children
4369 * have also been removed (modulo rounding error or bandwidth
4370 * control); however, such cases are rare and we can fix these
4373 * TODO: fix up out-of-order children on enqueue.
4375 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4376 list_del_leaf_cfs_rq(cfs_rq);
4378 struct rq *rq = rq_of(cfs_rq);
4379 update_rq_runnable_avg(rq, rq->nr_running);
4383 static void update_blocked_averages(int cpu)
4385 struct rq *rq = cpu_rq(cpu);
4386 struct cfs_rq *cfs_rq;
4387 unsigned long flags;
4389 raw_spin_lock_irqsave(&rq->lock, flags);
4390 update_rq_clock(rq);
4392 * Iterates the task_group tree in a bottom up fashion, see
4393 * list_add_leaf_cfs_rq() for details.
4395 for_each_leaf_cfs_rq(rq, cfs_rq) {
4397 * Note: We may want to consider periodically releasing
4398 * rq->lock about these updates so that creating many task
4399 * groups does not result in continually extending hold time.
4401 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4404 raw_spin_unlock_irqrestore(&rq->lock, flags);
4408 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4409 * This needs to be done in a top-down fashion because the load of a child
4410 * group is a fraction of its parents load.
4412 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4414 struct rq *rq = rq_of(cfs_rq);
4415 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4416 unsigned long now = jiffies;
4419 if (cfs_rq->last_h_load_update == now)
4422 cfs_rq->h_load_next = NULL;
4423 for_each_sched_entity(se) {
4424 cfs_rq = cfs_rq_of(se);
4425 cfs_rq->h_load_next = se;
4426 if (cfs_rq->last_h_load_update == now)
4431 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4432 cfs_rq->last_h_load_update = now;
4435 while ((se = cfs_rq->h_load_next) != NULL) {
4436 load = cfs_rq->h_load;
4437 load = div64_ul(load * se->avg.load_avg_contrib,
4438 cfs_rq->runnable_load_avg + 1);
4439 cfs_rq = group_cfs_rq(se);
4440 cfs_rq->h_load = load;
4441 cfs_rq->last_h_load_update = now;
4445 static unsigned long task_h_load(struct task_struct *p)
4447 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4449 update_cfs_rq_h_load(cfs_rq);
4450 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4451 cfs_rq->runnable_load_avg + 1);
4454 static inline void update_blocked_averages(int cpu)
4458 static unsigned long task_h_load(struct task_struct *p)
4460 return p->se.avg.load_avg_contrib;
4464 /********** Helpers for find_busiest_group ************************/
4466 * sg_lb_stats - stats of a sched_group required for load_balancing
4468 struct sg_lb_stats {
4469 unsigned long avg_load; /*Avg load across the CPUs of the group */
4470 unsigned long group_load; /* Total load over the CPUs of the group */
4471 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4472 unsigned long load_per_task;
4473 unsigned long group_power;
4474 unsigned int sum_nr_running; /* Nr tasks running in the group */
4475 unsigned int group_capacity;
4476 unsigned int idle_cpus;
4477 unsigned int group_weight;
4478 int group_imb; /* Is there an imbalance in the group ? */
4479 int group_has_capacity; /* Is there extra capacity in the group? */
4483 * sd_lb_stats - Structure to store the statistics of a sched_domain
4484 * during load balancing.
4486 struct sd_lb_stats {
4487 struct sched_group *busiest; /* Busiest group in this sd */
4488 struct sched_group *local; /* Local group in this sd */
4489 unsigned long total_load; /* Total load of all groups in sd */
4490 unsigned long total_pwr; /* Total power of all groups in sd */
4491 unsigned long avg_load; /* Average load across all groups in sd */
4493 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4494 struct sg_lb_stats local_stat; /* Statistics of the local group */
4497 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4500 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4501 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4502 * We must however clear busiest_stat::avg_load because
4503 * update_sd_pick_busiest() reads this before assignment.
4505 *sds = (struct sd_lb_stats){
4517 * get_sd_load_idx - Obtain the load index for a given sched domain.
4518 * @sd: The sched_domain whose load_idx is to be obtained.
4519 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4521 * Return: The load index.
4523 static inline int get_sd_load_idx(struct sched_domain *sd,
4524 enum cpu_idle_type idle)
4530 load_idx = sd->busy_idx;
4533 case CPU_NEWLY_IDLE:
4534 load_idx = sd->newidle_idx;
4537 load_idx = sd->idle_idx;
4544 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4546 return SCHED_POWER_SCALE;
4549 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4551 return default_scale_freq_power(sd, cpu);
4554 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4556 unsigned long weight = sd->span_weight;
4557 unsigned long smt_gain = sd->smt_gain;
4564 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4566 return default_scale_smt_power(sd, cpu);
4569 static unsigned long scale_rt_power(int cpu)
4571 struct rq *rq = cpu_rq(cpu);
4572 u64 total, available, age_stamp, avg;
4575 * Since we're reading these variables without serialization make sure
4576 * we read them once before doing sanity checks on them.
4578 age_stamp = ACCESS_ONCE(rq->age_stamp);
4579 avg = ACCESS_ONCE(rq->rt_avg);
4581 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4583 if (unlikely(total < avg)) {
4584 /* Ensures that power won't end up being negative */
4587 available = total - avg;
4590 if (unlikely((s64)total < SCHED_POWER_SCALE))
4591 total = SCHED_POWER_SCALE;
4593 total >>= SCHED_POWER_SHIFT;
4595 return div_u64(available, total);
4598 static void update_cpu_power(struct sched_domain *sd, int cpu)
4600 unsigned long weight = sd->span_weight;
4601 unsigned long power = SCHED_POWER_SCALE;
4602 struct sched_group *sdg = sd->groups;
4604 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4605 if (sched_feat(ARCH_POWER))
4606 power *= arch_scale_smt_power(sd, cpu);
4608 power *= default_scale_smt_power(sd, cpu);
4610 power >>= SCHED_POWER_SHIFT;
4613 sdg->sgp->power_orig = power;
4615 if (sched_feat(ARCH_POWER))
4616 power *= arch_scale_freq_power(sd, cpu);
4618 power *= default_scale_freq_power(sd, cpu);
4620 power >>= SCHED_POWER_SHIFT;
4622 power *= scale_rt_power(cpu);
4623 power >>= SCHED_POWER_SHIFT;
4628 cpu_rq(cpu)->cpu_power = power;
4629 sdg->sgp->power = power;
4632 void update_group_power(struct sched_domain *sd, int cpu)
4634 struct sched_domain *child = sd->child;
4635 struct sched_group *group, *sdg = sd->groups;
4636 unsigned long power, power_orig;
4637 unsigned long interval;
4639 interval = msecs_to_jiffies(sd->balance_interval);
4640 interval = clamp(interval, 1UL, max_load_balance_interval);
4641 sdg->sgp->next_update = jiffies + interval;
4644 update_cpu_power(sd, cpu);
4648 power_orig = power = 0;
4650 if (child->flags & SD_OVERLAP) {
4652 * SD_OVERLAP domains cannot assume that child groups
4653 * span the current group.
4656 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4657 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4659 power_orig += sg->sgp->power_orig;
4660 power += sg->sgp->power;
4664 * !SD_OVERLAP domains can assume that child groups
4665 * span the current group.
4668 group = child->groups;
4670 power_orig += group->sgp->power_orig;
4671 power += group->sgp->power;
4672 group = group->next;
4673 } while (group != child->groups);
4676 sdg->sgp->power_orig = power_orig;
4677 sdg->sgp->power = power;
4681 * Try and fix up capacity for tiny siblings, this is needed when
4682 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4683 * which on its own isn't powerful enough.
4685 * See update_sd_pick_busiest() and check_asym_packing().
4688 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4691 * Only siblings can have significantly less than SCHED_POWER_SCALE
4693 if (!(sd->flags & SD_SHARE_CPUPOWER))
4697 * If ~90% of the cpu_power is still there, we're good.
4699 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4706 * Group imbalance indicates (and tries to solve) the problem where balancing
4707 * groups is inadequate due to tsk_cpus_allowed() constraints.
4709 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4710 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4713 * { 0 1 2 3 } { 4 5 6 7 }
4716 * If we were to balance group-wise we'd place two tasks in the first group and
4717 * two tasks in the second group. Clearly this is undesired as it will overload
4718 * cpu 3 and leave one of the cpus in the second group unused.
4720 * The current solution to this issue is detecting the skew in the first group
4721 * by noticing the lower domain failed to reach balance and had difficulty
4722 * moving tasks due to affinity constraints.
4724 * When this is so detected; this group becomes a candidate for busiest; see
4725 * update_sd_pick_busiest(). And calculcate_imbalance() and
4726 * find_busiest_group() avoid some of the usual balance conditions to allow it
4727 * to create an effective group imbalance.
4729 * This is a somewhat tricky proposition since the next run might not find the
4730 * group imbalance and decide the groups need to be balanced again. A most
4731 * subtle and fragile situation.
4734 static inline int sg_imbalanced(struct sched_group *group)
4736 return group->sgp->imbalance;
4740 * Compute the group capacity.
4742 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4743 * first dividing out the smt factor and computing the actual number of cores
4744 * and limit power unit capacity with that.
4746 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4748 unsigned int capacity, smt, cpus;
4749 unsigned int power, power_orig;
4751 power = group->sgp->power;
4752 power_orig = group->sgp->power_orig;
4753 cpus = group->group_weight;
4755 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4756 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4757 capacity = cpus / smt; /* cores */
4759 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4761 capacity = fix_small_capacity(env->sd, group);
4767 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4768 * @env: The load balancing environment.
4769 * @group: sched_group whose statistics are to be updated.
4770 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4771 * @local_group: Does group contain this_cpu.
4772 * @sgs: variable to hold the statistics for this group.
4774 static inline void update_sg_lb_stats(struct lb_env *env,
4775 struct sched_group *group, int load_idx,
4776 int local_group, struct sg_lb_stats *sgs)
4778 unsigned long nr_running;
4782 memset(sgs, 0, sizeof(*sgs));
4784 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4785 struct rq *rq = cpu_rq(i);
4787 nr_running = rq->nr_running;
4789 /* Bias balancing toward cpus of our domain */
4791 load = target_load(i, load_idx);
4793 load = source_load(i, load_idx);
4795 sgs->group_load += load;
4796 sgs->sum_nr_running += nr_running;
4797 sgs->sum_weighted_load += weighted_cpuload(i);
4802 /* Adjust by relative CPU power of the group */
4803 sgs->group_power = group->sgp->power;
4804 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4806 if (sgs->sum_nr_running)
4807 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4809 sgs->group_weight = group->group_weight;
4811 sgs->group_imb = sg_imbalanced(group);
4812 sgs->group_capacity = sg_capacity(env, group);
4814 if (sgs->group_capacity > sgs->sum_nr_running)
4815 sgs->group_has_capacity = 1;
4819 * update_sd_pick_busiest - return 1 on busiest group
4820 * @env: The load balancing environment.
4821 * @sds: sched_domain statistics
4822 * @sg: sched_group candidate to be checked for being the busiest
4823 * @sgs: sched_group statistics
4825 * Determine if @sg is a busier group than the previously selected
4828 * Return: %true if @sg is a busier group than the previously selected
4829 * busiest group. %false otherwise.
4831 static bool update_sd_pick_busiest(struct lb_env *env,
4832 struct sd_lb_stats *sds,
4833 struct sched_group *sg,
4834 struct sg_lb_stats *sgs)
4836 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4839 if (sgs->sum_nr_running > sgs->group_capacity)
4846 * ASYM_PACKING needs to move all the work to the lowest
4847 * numbered CPUs in the group, therefore mark all groups
4848 * higher than ourself as busy.
4850 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4851 env->dst_cpu < group_first_cpu(sg)) {
4855 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4863 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4864 * @env: The load balancing environment.
4865 * @balance: Should we balance.
4866 * @sds: variable to hold the statistics for this sched_domain.
4868 static inline void update_sd_lb_stats(struct lb_env *env,
4869 struct sd_lb_stats *sds)
4871 struct sched_domain *child = env->sd->child;
4872 struct sched_group *sg = env->sd->groups;
4873 struct sg_lb_stats tmp_sgs;
4874 int load_idx, prefer_sibling = 0;
4876 if (child && child->flags & SD_PREFER_SIBLING)
4879 load_idx = get_sd_load_idx(env->sd, env->idle);
4882 struct sg_lb_stats *sgs = &tmp_sgs;
4885 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4888 sgs = &sds->local_stat;
4890 if (env->idle != CPU_NEWLY_IDLE ||
4891 time_after_eq(jiffies, sg->sgp->next_update))
4892 update_group_power(env->sd, env->dst_cpu);
4895 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4901 * In case the child domain prefers tasks go to siblings
4902 * first, lower the sg capacity to one so that we'll try
4903 * and move all the excess tasks away. We lower the capacity
4904 * of a group only if the local group has the capacity to fit
4905 * these excess tasks, i.e. nr_running < group_capacity. The
4906 * extra check prevents the case where you always pull from the
4907 * heaviest group when it is already under-utilized (possible
4908 * with a large weight task outweighs the tasks on the system).
4910 if (prefer_sibling && sds->local &&
4911 sds->local_stat.group_has_capacity)
4912 sgs->group_capacity = min(sgs->group_capacity, 1U);
4914 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4916 sds->busiest_stat = *sgs;
4920 /* Now, start updating sd_lb_stats */
4921 sds->total_load += sgs->group_load;
4922 sds->total_pwr += sgs->group_power;
4925 } while (sg != env->sd->groups);
4929 * check_asym_packing - Check to see if the group is packed into the
4932 * This is primarily intended to used at the sibling level. Some
4933 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4934 * case of POWER7, it can move to lower SMT modes only when higher
4935 * threads are idle. When in lower SMT modes, the threads will
4936 * perform better since they share less core resources. Hence when we
4937 * have idle threads, we want them to be the higher ones.
4939 * This packing function is run on idle threads. It checks to see if
4940 * the busiest CPU in this domain (core in the P7 case) has a higher
4941 * CPU number than the packing function is being run on. Here we are
4942 * assuming lower CPU number will be equivalent to lower a SMT thread
4945 * Return: 1 when packing is required and a task should be moved to
4946 * this CPU. The amount of the imbalance is returned in *imbalance.
4948 * @env: The load balancing environment.
4949 * @sds: Statistics of the sched_domain which is to be packed
4951 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4955 if (!(env->sd->flags & SD_ASYM_PACKING))
4961 busiest_cpu = group_first_cpu(sds->busiest);
4962 if (env->dst_cpu > busiest_cpu)
4965 env->imbalance = DIV_ROUND_CLOSEST(
4966 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
4973 * fix_small_imbalance - Calculate the minor imbalance that exists
4974 * amongst the groups of a sched_domain, during
4976 * @env: The load balancing environment.
4977 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4980 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4982 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4983 unsigned int imbn = 2;
4984 unsigned long scaled_busy_load_per_task;
4985 struct sg_lb_stats *local, *busiest;
4987 local = &sds->local_stat;
4988 busiest = &sds->busiest_stat;
4990 if (!local->sum_nr_running)
4991 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
4992 else if (busiest->load_per_task > local->load_per_task)
4995 scaled_busy_load_per_task =
4996 (busiest->load_per_task * SCHED_POWER_SCALE) /
4997 busiest->group_power;
4999 if (busiest->avg_load + scaled_busy_load_per_task >=
5000 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5001 env->imbalance = busiest->load_per_task;
5006 * OK, we don't have enough imbalance to justify moving tasks,
5007 * however we may be able to increase total CPU power used by
5011 pwr_now += busiest->group_power *
5012 min(busiest->load_per_task, busiest->avg_load);
5013 pwr_now += local->group_power *
5014 min(local->load_per_task, local->avg_load);
5015 pwr_now /= SCHED_POWER_SCALE;
5017 /* Amount of load we'd subtract */
5018 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5019 busiest->group_power;
5020 if (busiest->avg_load > tmp) {
5021 pwr_move += busiest->group_power *
5022 min(busiest->load_per_task,
5023 busiest->avg_load - tmp);
5026 /* Amount of load we'd add */
5027 if (busiest->avg_load * busiest->group_power <
5028 busiest->load_per_task * SCHED_POWER_SCALE) {
5029 tmp = (busiest->avg_load * busiest->group_power) /
5032 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5035 pwr_move += local->group_power *
5036 min(local->load_per_task, local->avg_load + tmp);
5037 pwr_move /= SCHED_POWER_SCALE;
5039 /* Move if we gain throughput */
5040 if (pwr_move > pwr_now)
5041 env->imbalance = busiest->load_per_task;
5045 * calculate_imbalance - Calculate the amount of imbalance present within the
5046 * groups of a given sched_domain during load balance.
5047 * @env: load balance environment
5048 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5050 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5052 unsigned long max_pull, load_above_capacity = ~0UL;
5053 struct sg_lb_stats *local, *busiest;
5055 local = &sds->local_stat;
5056 busiest = &sds->busiest_stat;
5058 if (busiest->group_imb) {
5060 * In the group_imb case we cannot rely on group-wide averages
5061 * to ensure cpu-load equilibrium, look at wider averages. XXX
5063 busiest->load_per_task =
5064 min(busiest->load_per_task, sds->avg_load);
5068 * In the presence of smp nice balancing, certain scenarios can have
5069 * max load less than avg load(as we skip the groups at or below
5070 * its cpu_power, while calculating max_load..)
5072 if (busiest->avg_load <= sds->avg_load ||
5073 local->avg_load >= sds->avg_load) {
5075 return fix_small_imbalance(env, sds);
5078 if (!busiest->group_imb) {
5080 * Don't want to pull so many tasks that a group would go idle.
5081 * Except of course for the group_imb case, since then we might
5082 * have to drop below capacity to reach cpu-load equilibrium.
5084 load_above_capacity =
5085 (busiest->sum_nr_running - busiest->group_capacity);
5087 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5088 load_above_capacity /= busiest->group_power;
5092 * We're trying to get all the cpus to the average_load, so we don't
5093 * want to push ourselves above the average load, nor do we wish to
5094 * reduce the max loaded cpu below the average load. At the same time,
5095 * we also don't want to reduce the group load below the group capacity
5096 * (so that we can implement power-savings policies etc). Thus we look
5097 * for the minimum possible imbalance.
5099 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5101 /* How much load to actually move to equalise the imbalance */
5102 env->imbalance = min(
5103 max_pull * busiest->group_power,
5104 (sds->avg_load - local->avg_load) * local->group_power
5105 ) / SCHED_POWER_SCALE;
5108 * if *imbalance is less than the average load per runnable task
5109 * there is no guarantee that any tasks will be moved so we'll have
5110 * a think about bumping its value to force at least one task to be
5113 if (env->imbalance < busiest->load_per_task)
5114 return fix_small_imbalance(env, sds);
5117 /******* find_busiest_group() helpers end here *********************/
5120 * find_busiest_group - Returns the busiest group within the sched_domain
5121 * if there is an imbalance. If there isn't an imbalance, and
5122 * the user has opted for power-savings, it returns a group whose
5123 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5124 * such a group exists.
5126 * Also calculates the amount of weighted load which should be moved
5127 * to restore balance.
5129 * @env: The load balancing environment.
5131 * Return: - The busiest group if imbalance exists.
5132 * - If no imbalance and user has opted for power-savings balance,
5133 * return the least loaded group whose CPUs can be
5134 * put to idle by rebalancing its tasks onto our group.
5136 static struct sched_group *find_busiest_group(struct lb_env *env)
5138 struct sg_lb_stats *local, *busiest;
5139 struct sd_lb_stats sds;
5141 init_sd_lb_stats(&sds);
5144 * Compute the various statistics relavent for load balancing at
5147 update_sd_lb_stats(env, &sds);
5148 local = &sds.local_stat;
5149 busiest = &sds.busiest_stat;
5151 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5152 check_asym_packing(env, &sds))
5155 /* There is no busy sibling group to pull tasks from */
5156 if (!sds.busiest || busiest->sum_nr_running == 0)
5159 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5162 * If the busiest group is imbalanced the below checks don't
5163 * work because they assume all things are equal, which typically
5164 * isn't true due to cpus_allowed constraints and the like.
5166 if (busiest->group_imb)
5169 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5170 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5171 !busiest->group_has_capacity)
5175 * If the local group is more busy than the selected busiest group
5176 * don't try and pull any tasks.
5178 if (local->avg_load >= busiest->avg_load)
5182 * Don't pull any tasks if this group is already above the domain
5185 if (local->avg_load >= sds.avg_load)
5188 if (env->idle == CPU_IDLE) {
5190 * This cpu is idle. If the busiest group load doesn't
5191 * have more tasks than the number of available cpu's and
5192 * there is no imbalance between this and busiest group
5193 * wrt to idle cpu's, it is balanced.
5195 if ((local->idle_cpus < busiest->idle_cpus) &&
5196 busiest->sum_nr_running <= busiest->group_weight)
5200 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5201 * imbalance_pct to be conservative.
5203 if (100 * busiest->avg_load <=
5204 env->sd->imbalance_pct * local->avg_load)
5209 /* Looks like there is an imbalance. Compute it */
5210 calculate_imbalance(env, &sds);
5219 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5221 static struct rq *find_busiest_queue(struct lb_env *env,
5222 struct sched_group *group)
5224 struct rq *busiest = NULL, *rq;
5225 unsigned long busiest_load = 0, busiest_power = 1;
5228 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5229 unsigned long power = power_of(i);
5230 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5235 capacity = fix_small_capacity(env->sd, group);
5238 wl = weighted_cpuload(i);
5241 * When comparing with imbalance, use weighted_cpuload()
5242 * which is not scaled with the cpu power.
5244 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5248 * For the load comparisons with the other cpu's, consider
5249 * the weighted_cpuload() scaled with the cpu power, so that
5250 * the load can be moved away from the cpu that is potentially
5251 * running at a lower capacity.
5253 * Thus we're looking for max(wl_i / power_i), crosswise
5254 * multiplication to rid ourselves of the division works out
5255 * to: wl_i * power_j > wl_j * power_i; where j is our
5258 if (wl * busiest_power > busiest_load * power) {
5260 busiest_power = power;
5269 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5270 * so long as it is large enough.
5272 #define MAX_PINNED_INTERVAL 512
5274 /* Working cpumask for load_balance and load_balance_newidle. */
5275 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5277 static int need_active_balance(struct lb_env *env)
5279 struct sched_domain *sd = env->sd;
5281 if (env->idle == CPU_NEWLY_IDLE) {
5284 * ASYM_PACKING needs to force migrate tasks from busy but
5285 * higher numbered CPUs in order to pack all tasks in the
5286 * lowest numbered CPUs.
5288 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5292 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5295 static int active_load_balance_cpu_stop(void *data);
5297 static int should_we_balance(struct lb_env *env)
5299 struct sched_group *sg = env->sd->groups;
5300 struct cpumask *sg_cpus, *sg_mask;
5301 int cpu, balance_cpu = -1;
5304 * In the newly idle case, we will allow all the cpu's
5305 * to do the newly idle load balance.
5307 if (env->idle == CPU_NEWLY_IDLE)
5310 sg_cpus = sched_group_cpus(sg);
5311 sg_mask = sched_group_mask(sg);
5312 /* Try to find first idle cpu */
5313 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5314 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5321 if (balance_cpu == -1)
5322 balance_cpu = group_balance_cpu(sg);
5325 * First idle cpu or the first cpu(busiest) in this sched group
5326 * is eligible for doing load balancing at this and above domains.
5328 return balance_cpu == env->dst_cpu;
5332 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5333 * tasks if there is an imbalance.
5335 static int load_balance(int this_cpu, struct rq *this_rq,
5336 struct sched_domain *sd, enum cpu_idle_type idle,
5337 int *continue_balancing)
5339 int ld_moved, cur_ld_moved, active_balance = 0;
5340 struct sched_domain *sd_parent = sd->parent;
5341 struct sched_group *group;
5343 unsigned long flags;
5344 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5346 struct lb_env env = {
5348 .dst_cpu = this_cpu,
5350 .dst_grpmask = sched_group_cpus(sd->groups),
5352 .loop_break = sched_nr_migrate_break,
5357 * For NEWLY_IDLE load_balancing, we don't need to consider
5358 * other cpus in our group
5360 if (idle == CPU_NEWLY_IDLE)
5361 env.dst_grpmask = NULL;
5363 cpumask_copy(cpus, cpu_active_mask);
5365 schedstat_inc(sd, lb_count[idle]);
5368 if (!should_we_balance(&env)) {
5369 *continue_balancing = 0;
5373 group = find_busiest_group(&env);
5375 schedstat_inc(sd, lb_nobusyg[idle]);
5379 busiest = find_busiest_queue(&env, group);
5381 schedstat_inc(sd, lb_nobusyq[idle]);
5385 BUG_ON(busiest == env.dst_rq);
5387 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5390 if (busiest->nr_running > 1) {
5392 * Attempt to move tasks. If find_busiest_group has found
5393 * an imbalance but busiest->nr_running <= 1, the group is
5394 * still unbalanced. ld_moved simply stays zero, so it is
5395 * correctly treated as an imbalance.
5397 env.flags |= LBF_ALL_PINNED;
5398 env.src_cpu = busiest->cpu;
5399 env.src_rq = busiest;
5400 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5403 local_irq_save(flags);
5404 double_rq_lock(env.dst_rq, busiest);
5407 * cur_ld_moved - load moved in current iteration
5408 * ld_moved - cumulative load moved across iterations
5410 cur_ld_moved = move_tasks(&env);
5411 ld_moved += cur_ld_moved;
5412 double_rq_unlock(env.dst_rq, busiest);
5413 local_irq_restore(flags);
5416 * some other cpu did the load balance for us.
5418 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5419 resched_cpu(env.dst_cpu);
5421 if (env.flags & LBF_NEED_BREAK) {
5422 env.flags &= ~LBF_NEED_BREAK;
5427 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5428 * us and move them to an alternate dst_cpu in our sched_group
5429 * where they can run. The upper limit on how many times we
5430 * iterate on same src_cpu is dependent on number of cpus in our
5433 * This changes load balance semantics a bit on who can move
5434 * load to a given_cpu. In addition to the given_cpu itself
5435 * (or a ilb_cpu acting on its behalf where given_cpu is
5436 * nohz-idle), we now have balance_cpu in a position to move
5437 * load to given_cpu. In rare situations, this may cause
5438 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5439 * _independently_ and at _same_ time to move some load to
5440 * given_cpu) causing exceess load to be moved to given_cpu.
5441 * This however should not happen so much in practice and
5442 * moreover subsequent load balance cycles should correct the
5443 * excess load moved.
5445 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5447 /* Prevent to re-select dst_cpu via env's cpus */
5448 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5450 env.dst_rq = cpu_rq(env.new_dst_cpu);
5451 env.dst_cpu = env.new_dst_cpu;
5452 env.flags &= ~LBF_DST_PINNED;
5454 env.loop_break = sched_nr_migrate_break;
5457 * Go back to "more_balance" rather than "redo" since we
5458 * need to continue with same src_cpu.
5464 * We failed to reach balance because of affinity.
5467 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5469 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5470 *group_imbalance = 1;
5471 } else if (*group_imbalance)
5472 *group_imbalance = 0;
5475 /* All tasks on this runqueue were pinned by CPU affinity */
5476 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5477 cpumask_clear_cpu(cpu_of(busiest), cpus);
5478 if (!cpumask_empty(cpus)) {
5480 env.loop_break = sched_nr_migrate_break;
5488 schedstat_inc(sd, lb_failed[idle]);
5490 * Increment the failure counter only on periodic balance.
5491 * We do not want newidle balance, which can be very
5492 * frequent, pollute the failure counter causing
5493 * excessive cache_hot migrations and active balances.
5495 if (idle != CPU_NEWLY_IDLE)
5496 sd->nr_balance_failed++;
5498 if (need_active_balance(&env)) {
5499 raw_spin_lock_irqsave(&busiest->lock, flags);
5501 /* don't kick the active_load_balance_cpu_stop,
5502 * if the curr task on busiest cpu can't be
5505 if (!cpumask_test_cpu(this_cpu,
5506 tsk_cpus_allowed(busiest->curr))) {
5507 raw_spin_unlock_irqrestore(&busiest->lock,
5509 env.flags |= LBF_ALL_PINNED;
5510 goto out_one_pinned;
5514 * ->active_balance synchronizes accesses to
5515 * ->active_balance_work. Once set, it's cleared
5516 * only after active load balance is finished.
5518 if (!busiest->active_balance) {
5519 busiest->active_balance = 1;
5520 busiest->push_cpu = this_cpu;
5523 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5525 if (active_balance) {
5526 stop_one_cpu_nowait(cpu_of(busiest),
5527 active_load_balance_cpu_stop, busiest,
5528 &busiest->active_balance_work);
5532 * We've kicked active balancing, reset the failure
5535 sd->nr_balance_failed = sd->cache_nice_tries+1;
5538 sd->nr_balance_failed = 0;
5540 if (likely(!active_balance)) {
5541 /* We were unbalanced, so reset the balancing interval */
5542 sd->balance_interval = sd->min_interval;
5545 * If we've begun active balancing, start to back off. This
5546 * case may not be covered by the all_pinned logic if there
5547 * is only 1 task on the busy runqueue (because we don't call
5550 if (sd->balance_interval < sd->max_interval)
5551 sd->balance_interval *= 2;
5557 schedstat_inc(sd, lb_balanced[idle]);
5559 sd->nr_balance_failed = 0;
5562 /* tune up the balancing interval */
5563 if (((env.flags & LBF_ALL_PINNED) &&
5564 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5565 (sd->balance_interval < sd->max_interval))
5566 sd->balance_interval *= 2;
5574 * idle_balance is called by schedule() if this_cpu is about to become
5575 * idle. Attempts to pull tasks from other CPUs.
5577 void idle_balance(int this_cpu, struct rq *this_rq)
5579 struct sched_domain *sd;
5580 int pulled_task = 0;
5581 unsigned long next_balance = jiffies + HZ;
5584 this_rq->idle_stamp = rq_clock(this_rq);
5586 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5590 * Drop the rq->lock, but keep IRQ/preempt disabled.
5592 raw_spin_unlock(&this_rq->lock);
5594 update_blocked_averages(this_cpu);
5596 for_each_domain(this_cpu, sd) {
5597 unsigned long interval;
5598 int continue_balancing = 1;
5599 u64 t0, domain_cost;
5601 if (!(sd->flags & SD_LOAD_BALANCE))
5604 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5607 if (sd->flags & SD_BALANCE_NEWIDLE) {
5608 t0 = sched_clock_cpu(this_cpu);
5610 /* If we've pulled tasks over stop searching: */
5611 pulled_task = load_balance(this_cpu, this_rq,
5613 &continue_balancing);
5615 domain_cost = sched_clock_cpu(this_cpu) - t0;
5616 if (domain_cost > sd->max_newidle_lb_cost)
5617 sd->max_newidle_lb_cost = domain_cost;
5619 curr_cost += domain_cost;
5622 interval = msecs_to_jiffies(sd->balance_interval);
5623 if (time_after(next_balance, sd->last_balance + interval))
5624 next_balance = sd->last_balance + interval;
5626 this_rq->idle_stamp = 0;
5632 raw_spin_lock(&this_rq->lock);
5634 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5636 * We are going idle. next_balance may be set based on
5637 * a busy processor. So reset next_balance.
5639 this_rq->next_balance = next_balance;
5642 if (curr_cost > this_rq->max_idle_balance_cost)
5643 this_rq->max_idle_balance_cost = curr_cost;
5647 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5648 * running tasks off the busiest CPU onto idle CPUs. It requires at
5649 * least 1 task to be running on each physical CPU where possible, and
5650 * avoids physical / logical imbalances.
5652 static int active_load_balance_cpu_stop(void *data)
5654 struct rq *busiest_rq = data;
5655 int busiest_cpu = cpu_of(busiest_rq);
5656 int target_cpu = busiest_rq->push_cpu;
5657 struct rq *target_rq = cpu_rq(target_cpu);
5658 struct sched_domain *sd;
5660 raw_spin_lock_irq(&busiest_rq->lock);
5662 /* make sure the requested cpu hasn't gone down in the meantime */
5663 if (unlikely(busiest_cpu != smp_processor_id() ||
5664 !busiest_rq->active_balance))
5667 /* Is there any task to move? */
5668 if (busiest_rq->nr_running <= 1)
5672 * This condition is "impossible", if it occurs
5673 * we need to fix it. Originally reported by
5674 * Bjorn Helgaas on a 128-cpu setup.
5676 BUG_ON(busiest_rq == target_rq);
5678 /* move a task from busiest_rq to target_rq */
5679 double_lock_balance(busiest_rq, target_rq);
5681 /* Search for an sd spanning us and the target CPU. */
5683 for_each_domain(target_cpu, sd) {
5684 if ((sd->flags & SD_LOAD_BALANCE) &&
5685 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5690 struct lb_env env = {
5692 .dst_cpu = target_cpu,
5693 .dst_rq = target_rq,
5694 .src_cpu = busiest_rq->cpu,
5695 .src_rq = busiest_rq,
5699 schedstat_inc(sd, alb_count);
5701 if (move_one_task(&env))
5702 schedstat_inc(sd, alb_pushed);
5704 schedstat_inc(sd, alb_failed);
5707 double_unlock_balance(busiest_rq, target_rq);
5709 busiest_rq->active_balance = 0;
5710 raw_spin_unlock_irq(&busiest_rq->lock);
5714 #ifdef CONFIG_NO_HZ_COMMON
5716 * idle load balancing details
5717 * - When one of the busy CPUs notice that there may be an idle rebalancing
5718 * needed, they will kick the idle load balancer, which then does idle
5719 * load balancing for all the idle CPUs.
5722 cpumask_var_t idle_cpus_mask;
5724 unsigned long next_balance; /* in jiffy units */
5725 } nohz ____cacheline_aligned;
5727 static inline int find_new_ilb(int call_cpu)
5729 int ilb = cpumask_first(nohz.idle_cpus_mask);
5731 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5738 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5739 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5740 * CPU (if there is one).
5742 static void nohz_balancer_kick(int cpu)
5746 nohz.next_balance++;
5748 ilb_cpu = find_new_ilb(cpu);
5750 if (ilb_cpu >= nr_cpu_ids)
5753 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5756 * Use smp_send_reschedule() instead of resched_cpu().
5757 * This way we generate a sched IPI on the target cpu which
5758 * is idle. And the softirq performing nohz idle load balance
5759 * will be run before returning from the IPI.
5761 smp_send_reschedule(ilb_cpu);
5765 static inline void nohz_balance_exit_idle(int cpu)
5767 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5768 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5769 atomic_dec(&nohz.nr_cpus);
5770 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5774 static inline void set_cpu_sd_state_busy(void)
5776 struct sched_domain *sd;
5779 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5781 if (!sd || !sd->nohz_idle)
5785 for (; sd; sd = sd->parent)
5786 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5791 void set_cpu_sd_state_idle(void)
5793 struct sched_domain *sd;
5796 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5798 if (!sd || sd->nohz_idle)
5802 for (; sd; sd = sd->parent)
5803 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5809 * This routine will record that the cpu is going idle with tick stopped.
5810 * This info will be used in performing idle load balancing in the future.
5812 void nohz_balance_enter_idle(int cpu)
5815 * If this cpu is going down, then nothing needs to be done.
5817 if (!cpu_active(cpu))
5820 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5823 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5824 atomic_inc(&nohz.nr_cpus);
5825 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5828 static int sched_ilb_notifier(struct notifier_block *nfb,
5829 unsigned long action, void *hcpu)
5831 switch (action & ~CPU_TASKS_FROZEN) {
5833 nohz_balance_exit_idle(smp_processor_id());
5841 static DEFINE_SPINLOCK(balancing);
5844 * Scale the max load_balance interval with the number of CPUs in the system.
5845 * This trades load-balance latency on larger machines for less cross talk.
5847 void update_max_interval(void)
5849 max_load_balance_interval = HZ*num_online_cpus()/10;
5853 * It checks each scheduling domain to see if it is due to be balanced,
5854 * and initiates a balancing operation if so.
5856 * Balancing parameters are set up in init_sched_domains.
5858 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5860 int continue_balancing = 1;
5861 struct rq *rq = cpu_rq(cpu);
5862 unsigned long interval;
5863 struct sched_domain *sd;
5864 /* Earliest time when we have to do rebalance again */
5865 unsigned long next_balance = jiffies + 60*HZ;
5866 int update_next_balance = 0;
5867 int need_serialize, need_decay = 0;
5870 update_blocked_averages(cpu);
5873 for_each_domain(cpu, sd) {
5875 * Decay the newidle max times here because this is a regular
5876 * visit to all the domains. Decay ~1% per second.
5878 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5879 sd->max_newidle_lb_cost =
5880 (sd->max_newidle_lb_cost * 253) / 256;
5881 sd->next_decay_max_lb_cost = jiffies + HZ;
5884 max_cost += sd->max_newidle_lb_cost;
5886 if (!(sd->flags & SD_LOAD_BALANCE))
5890 * Stop the load balance at this level. There is another
5891 * CPU in our sched group which is doing load balancing more
5894 if (!continue_balancing) {
5900 interval = sd->balance_interval;
5901 if (idle != CPU_IDLE)
5902 interval *= sd->busy_factor;
5904 /* scale ms to jiffies */
5905 interval = msecs_to_jiffies(interval);
5906 interval = clamp(interval, 1UL, max_load_balance_interval);
5908 need_serialize = sd->flags & SD_SERIALIZE;
5910 if (need_serialize) {
5911 if (!spin_trylock(&balancing))
5915 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5916 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5918 * The LBF_DST_PINNED logic could have changed
5919 * env->dst_cpu, so we can't know our idle
5920 * state even if we migrated tasks. Update it.
5922 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5924 sd->last_balance = jiffies;
5927 spin_unlock(&balancing);
5929 if (time_after(next_balance, sd->last_balance + interval)) {
5930 next_balance = sd->last_balance + interval;
5931 update_next_balance = 1;
5936 * Ensure the rq-wide value also decays but keep it at a
5937 * reasonable floor to avoid funnies with rq->avg_idle.
5939 rq->max_idle_balance_cost =
5940 max((u64)sysctl_sched_migration_cost, max_cost);
5945 * next_balance will be updated only when there is a need.
5946 * When the cpu is attached to null domain for ex, it will not be
5949 if (likely(update_next_balance))
5950 rq->next_balance = next_balance;
5953 #ifdef CONFIG_NO_HZ_COMMON
5955 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
5956 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5958 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5960 struct rq *this_rq = cpu_rq(this_cpu);
5964 if (idle != CPU_IDLE ||
5965 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5968 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5969 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5973 * If this cpu gets work to do, stop the load balancing
5974 * work being done for other cpus. Next load
5975 * balancing owner will pick it up.
5980 rq = cpu_rq(balance_cpu);
5982 raw_spin_lock_irq(&rq->lock);
5983 update_rq_clock(rq);
5984 update_idle_cpu_load(rq);
5985 raw_spin_unlock_irq(&rq->lock);
5987 rebalance_domains(balance_cpu, CPU_IDLE);
5989 if (time_after(this_rq->next_balance, rq->next_balance))
5990 this_rq->next_balance = rq->next_balance;
5992 nohz.next_balance = this_rq->next_balance;
5994 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5998 * Current heuristic for kicking the idle load balancer in the presence
5999 * of an idle cpu is the system.
6000 * - This rq has more than one task.
6001 * - At any scheduler domain level, this cpu's scheduler group has multiple
6002 * busy cpu's exceeding the group's power.
6003 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6004 * domain span are idle.
6006 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6008 unsigned long now = jiffies;
6009 struct sched_domain *sd;
6011 if (unlikely(idle_cpu(cpu)))
6015 * We may be recently in ticked or tickless idle mode. At the first
6016 * busy tick after returning from idle, we will update the busy stats.
6018 set_cpu_sd_state_busy();
6019 nohz_balance_exit_idle(cpu);
6022 * None are in tickless mode and hence no need for NOHZ idle load
6025 if (likely(!atomic_read(&nohz.nr_cpus)))
6028 if (time_before(now, nohz.next_balance))
6031 if (rq->nr_running >= 2)
6035 for_each_domain(cpu, sd) {
6036 struct sched_group *sg = sd->groups;
6037 struct sched_group_power *sgp = sg->sgp;
6038 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6040 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6041 goto need_kick_unlock;
6043 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6044 && (cpumask_first_and(nohz.idle_cpus_mask,
6045 sched_domain_span(sd)) < cpu))
6046 goto need_kick_unlock;
6048 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6060 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6064 * run_rebalance_domains is triggered when needed from the scheduler tick.
6065 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6067 static void run_rebalance_domains(struct softirq_action *h)
6069 int this_cpu = smp_processor_id();
6070 struct rq *this_rq = cpu_rq(this_cpu);
6071 enum cpu_idle_type idle = this_rq->idle_balance ?
6072 CPU_IDLE : CPU_NOT_IDLE;
6074 rebalance_domains(this_cpu, idle);
6077 * If this cpu has a pending nohz_balance_kick, then do the
6078 * balancing on behalf of the other idle cpus whose ticks are
6081 nohz_idle_balance(this_cpu, idle);
6084 static inline int on_null_domain(int cpu)
6086 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6090 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6092 void trigger_load_balance(struct rq *rq, int cpu)
6094 /* Don't need to rebalance while attached to NULL domain */
6095 if (time_after_eq(jiffies, rq->next_balance) &&
6096 likely(!on_null_domain(cpu)))
6097 raise_softirq(SCHED_SOFTIRQ);
6098 #ifdef CONFIG_NO_HZ_COMMON
6099 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6100 nohz_balancer_kick(cpu);
6104 static void rq_online_fair(struct rq *rq)
6109 static void rq_offline_fair(struct rq *rq)
6113 /* Ensure any throttled groups are reachable by pick_next_task */
6114 unthrottle_offline_cfs_rqs(rq);
6117 #endif /* CONFIG_SMP */
6120 * scheduler tick hitting a task of our scheduling class:
6122 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6124 struct cfs_rq *cfs_rq;
6125 struct sched_entity *se = &curr->se;
6127 for_each_sched_entity(se) {
6128 cfs_rq = cfs_rq_of(se);
6129 entity_tick(cfs_rq, se, queued);
6132 if (numabalancing_enabled)
6133 task_tick_numa(rq, curr);
6135 update_rq_runnable_avg(rq, 1);
6139 * called on fork with the child task as argument from the parent's context
6140 * - child not yet on the tasklist
6141 * - preemption disabled
6143 static void task_fork_fair(struct task_struct *p)
6145 struct cfs_rq *cfs_rq;
6146 struct sched_entity *se = &p->se, *curr;
6147 int this_cpu = smp_processor_id();
6148 struct rq *rq = this_rq();
6149 unsigned long flags;
6151 raw_spin_lock_irqsave(&rq->lock, flags);
6153 update_rq_clock(rq);
6155 cfs_rq = task_cfs_rq(current);
6156 curr = cfs_rq->curr;
6159 * Not only the cpu but also the task_group of the parent might have
6160 * been changed after parent->se.parent,cfs_rq were copied to
6161 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6162 * of child point to valid ones.
6165 __set_task_cpu(p, this_cpu);
6168 update_curr(cfs_rq);
6171 se->vruntime = curr->vruntime;
6172 place_entity(cfs_rq, se, 1);
6174 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6176 * Upon rescheduling, sched_class::put_prev_task() will place
6177 * 'current' within the tree based on its new key value.
6179 swap(curr->vruntime, se->vruntime);
6180 resched_task(rq->curr);
6183 se->vruntime -= cfs_rq->min_vruntime;
6185 raw_spin_unlock_irqrestore(&rq->lock, flags);
6189 * Priority of the task has changed. Check to see if we preempt
6193 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6199 * Reschedule if we are currently running on this runqueue and
6200 * our priority decreased, or if we are not currently running on
6201 * this runqueue and our priority is higher than the current's
6203 if (rq->curr == p) {
6204 if (p->prio > oldprio)
6205 resched_task(rq->curr);
6207 check_preempt_curr(rq, p, 0);
6210 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6212 struct sched_entity *se = &p->se;
6213 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6216 * Ensure the task's vruntime is normalized, so that when its
6217 * switched back to the fair class the enqueue_entity(.flags=0) will
6218 * do the right thing.
6220 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6221 * have normalized the vruntime, if it was !on_rq, then only when
6222 * the task is sleeping will it still have non-normalized vruntime.
6224 if (!se->on_rq && p->state != TASK_RUNNING) {
6226 * Fix up our vruntime so that the current sleep doesn't
6227 * cause 'unlimited' sleep bonus.
6229 place_entity(cfs_rq, se, 0);
6230 se->vruntime -= cfs_rq->min_vruntime;
6235 * Remove our load from contribution when we leave sched_fair
6236 * and ensure we don't carry in an old decay_count if we
6239 if (se->avg.decay_count) {
6240 __synchronize_entity_decay(se);
6241 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6247 * We switched to the sched_fair class.
6249 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6255 * We were most likely switched from sched_rt, so
6256 * kick off the schedule if running, otherwise just see
6257 * if we can still preempt the current task.
6260 resched_task(rq->curr);
6262 check_preempt_curr(rq, p, 0);
6265 /* Account for a task changing its policy or group.
6267 * This routine is mostly called to set cfs_rq->curr field when a task
6268 * migrates between groups/classes.
6270 static void set_curr_task_fair(struct rq *rq)
6272 struct sched_entity *se = &rq->curr->se;
6274 for_each_sched_entity(se) {
6275 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6277 set_next_entity(cfs_rq, se);
6278 /* ensure bandwidth has been allocated on our new cfs_rq */
6279 account_cfs_rq_runtime(cfs_rq, 0);
6283 void init_cfs_rq(struct cfs_rq *cfs_rq)
6285 cfs_rq->tasks_timeline = RB_ROOT;
6286 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6287 #ifndef CONFIG_64BIT
6288 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6291 atomic64_set(&cfs_rq->decay_counter, 1);
6292 atomic_long_set(&cfs_rq->removed_load, 0);
6296 #ifdef CONFIG_FAIR_GROUP_SCHED
6297 static void task_move_group_fair(struct task_struct *p, int on_rq)
6299 struct cfs_rq *cfs_rq;
6301 * If the task was not on the rq at the time of this cgroup movement
6302 * it must have been asleep, sleeping tasks keep their ->vruntime
6303 * absolute on their old rq until wakeup (needed for the fair sleeper
6304 * bonus in place_entity()).
6306 * If it was on the rq, we've just 'preempted' it, which does convert
6307 * ->vruntime to a relative base.
6309 * Make sure both cases convert their relative position when migrating
6310 * to another cgroup's rq. This does somewhat interfere with the
6311 * fair sleeper stuff for the first placement, but who cares.
6314 * When !on_rq, vruntime of the task has usually NOT been normalized.
6315 * But there are some cases where it has already been normalized:
6317 * - Moving a forked child which is waiting for being woken up by
6318 * wake_up_new_task().
6319 * - Moving a task which has been woken up by try_to_wake_up() and
6320 * waiting for actually being woken up by sched_ttwu_pending().
6322 * To prevent boost or penalty in the new cfs_rq caused by delta
6323 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6325 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6329 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6330 set_task_rq(p, task_cpu(p));
6332 cfs_rq = cfs_rq_of(&p->se);
6333 p->se.vruntime += cfs_rq->min_vruntime;
6336 * migrate_task_rq_fair() will have removed our previous
6337 * contribution, but we must synchronize for ongoing future
6340 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6341 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6346 void free_fair_sched_group(struct task_group *tg)
6350 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6352 for_each_possible_cpu(i) {
6354 kfree(tg->cfs_rq[i]);
6363 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6365 struct cfs_rq *cfs_rq;
6366 struct sched_entity *se;
6369 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6372 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6376 tg->shares = NICE_0_LOAD;
6378 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6380 for_each_possible_cpu(i) {
6381 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6382 GFP_KERNEL, cpu_to_node(i));
6386 se = kzalloc_node(sizeof(struct sched_entity),
6387 GFP_KERNEL, cpu_to_node(i));
6391 init_cfs_rq(cfs_rq);
6392 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6403 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6405 struct rq *rq = cpu_rq(cpu);
6406 unsigned long flags;
6409 * Only empty task groups can be destroyed; so we can speculatively
6410 * check on_list without danger of it being re-added.
6412 if (!tg->cfs_rq[cpu]->on_list)
6415 raw_spin_lock_irqsave(&rq->lock, flags);
6416 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6417 raw_spin_unlock_irqrestore(&rq->lock, flags);
6420 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6421 struct sched_entity *se, int cpu,
6422 struct sched_entity *parent)
6424 struct rq *rq = cpu_rq(cpu);
6428 init_cfs_rq_runtime(cfs_rq);
6430 tg->cfs_rq[cpu] = cfs_rq;
6433 /* se could be NULL for root_task_group */
6438 se->cfs_rq = &rq->cfs;
6440 se->cfs_rq = parent->my_q;
6443 update_load_set(&se->load, 0);
6444 se->parent = parent;
6447 static DEFINE_MUTEX(shares_mutex);
6449 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6452 unsigned long flags;
6455 * We can't change the weight of the root cgroup.
6460 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6462 mutex_lock(&shares_mutex);
6463 if (tg->shares == shares)
6466 tg->shares = shares;
6467 for_each_possible_cpu(i) {
6468 struct rq *rq = cpu_rq(i);
6469 struct sched_entity *se;
6472 /* Propagate contribution to hierarchy */
6473 raw_spin_lock_irqsave(&rq->lock, flags);
6475 /* Possible calls to update_curr() need rq clock */
6476 update_rq_clock(rq);
6477 for_each_sched_entity(se)
6478 update_cfs_shares(group_cfs_rq(se));
6479 raw_spin_unlock_irqrestore(&rq->lock, flags);
6483 mutex_unlock(&shares_mutex);
6486 #else /* CONFIG_FAIR_GROUP_SCHED */
6488 void free_fair_sched_group(struct task_group *tg) { }
6490 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6495 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6497 #endif /* CONFIG_FAIR_GROUP_SCHED */
6500 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6502 struct sched_entity *se = &task->se;
6503 unsigned int rr_interval = 0;
6506 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6509 if (rq->cfs.load.weight)
6510 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6516 * All the scheduling class methods:
6518 const struct sched_class fair_sched_class = {
6519 .next = &idle_sched_class,
6520 .enqueue_task = enqueue_task_fair,
6521 .dequeue_task = dequeue_task_fair,
6522 .yield_task = yield_task_fair,
6523 .yield_to_task = yield_to_task_fair,
6525 .check_preempt_curr = check_preempt_wakeup,
6527 .pick_next_task = pick_next_task_fair,
6528 .put_prev_task = put_prev_task_fair,
6531 .select_task_rq = select_task_rq_fair,
6532 .migrate_task_rq = migrate_task_rq_fair,
6534 .rq_online = rq_online_fair,
6535 .rq_offline = rq_offline_fair,
6537 .task_waking = task_waking_fair,
6540 .set_curr_task = set_curr_task_fair,
6541 .task_tick = task_tick_fair,
6542 .task_fork = task_fork_fair,
6544 .prio_changed = prio_changed_fair,
6545 .switched_from = switched_from_fair,
6546 .switched_to = switched_to_fair,
6548 .get_rr_interval = get_rr_interval_fair,
6550 #ifdef CONFIG_FAIR_GROUP_SCHED
6551 .task_move_group = task_move_group_fair,
6555 #ifdef CONFIG_SCHED_DEBUG
6556 void print_cfs_stats(struct seq_file *m, int cpu)
6558 struct cfs_rq *cfs_rq;
6561 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6562 print_cfs_rq(m, cpu, cfs_rq);
6567 __init void init_sched_fair_class(void)
6570 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6572 #ifdef CONFIG_NO_HZ_COMMON
6573 nohz.next_balance = jiffies;
6574 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6575 cpu_notifier(sched_ilb_notifier, 0);