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 unsigned long weighted_cpuload(const int cpu);
893 find_idlest_cpu_node(int this_cpu, int nid)
895 unsigned long load, min_load = ULONG_MAX;
896 int i, idlest_cpu = this_cpu;
898 BUG_ON(cpu_to_node(this_cpu) == nid);
901 for_each_cpu(i, cpumask_of_node(nid)) {
902 load = weighted_cpuload(i);
904 if (load < min_load) {
914 static void task_numa_placement(struct task_struct *p)
916 int seq, nid, max_nid = -1;
917 unsigned long max_faults = 0;
919 if (!p->mm) /* for example, ksmd faulting in a user's mm */
921 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
922 if (p->numa_scan_seq == seq)
924 p->numa_scan_seq = seq;
925 p->numa_migrate_seq++;
926 p->numa_scan_period_max = task_scan_max(p);
928 /* Find the node with the highest number of faults */
929 for_each_online_node(nid) {
930 unsigned long faults;
932 /* Decay existing window and copy faults since last scan */
933 p->numa_faults[nid] >>= 1;
934 p->numa_faults[nid] += p->numa_faults_buffer[nid];
935 p->numa_faults_buffer[nid] = 0;
937 faults = p->numa_faults[nid];
938 if (faults > max_faults) {
945 * Record the preferred node as the node with the most faults,
946 * requeue the task to be running on the idlest CPU on the
947 * preferred node and reset the scanning rate to recheck
948 * the working set placement.
950 if (max_faults && max_nid != p->numa_preferred_nid) {
954 * If the task is not on the preferred node then find the most
955 * idle CPU to migrate to.
957 preferred_cpu = task_cpu(p);
958 if (cpu_to_node(preferred_cpu) != max_nid) {
959 preferred_cpu = find_idlest_cpu_node(preferred_cpu,
963 /* Update the preferred nid and migrate task if possible */
964 p->numa_preferred_nid = max_nid;
965 p->numa_migrate_seq = 0;
966 migrate_task_to(p, preferred_cpu);
971 * Got a PROT_NONE fault for a page on @node.
973 void task_numa_fault(int node, int pages, bool migrated)
975 struct task_struct *p = current;
977 if (!numabalancing_enabled)
980 /* Allocate buffer to track faults on a per-node basis */
981 if (unlikely(!p->numa_faults)) {
982 int size = sizeof(*p->numa_faults) * nr_node_ids;
984 /* numa_faults and numa_faults_buffer share the allocation */
985 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
989 BUG_ON(p->numa_faults_buffer);
990 p->numa_faults_buffer = p->numa_faults + nr_node_ids;
994 * If pages are properly placed (did not migrate) then scan slower.
995 * This is reset periodically in case of phase changes
998 /* Initialise if necessary */
999 if (!p->numa_scan_period_max)
1000 p->numa_scan_period_max = task_scan_max(p);
1002 p->numa_scan_period = min(p->numa_scan_period_max,
1003 p->numa_scan_period + 10);
1006 task_numa_placement(p);
1008 p->numa_faults_buffer[node] += pages;
1011 static void reset_ptenuma_scan(struct task_struct *p)
1013 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1014 p->mm->numa_scan_offset = 0;
1018 * The expensive part of numa migration is done from task_work context.
1019 * Triggered from task_tick_numa().
1021 void task_numa_work(struct callback_head *work)
1023 unsigned long migrate, next_scan, now = jiffies;
1024 struct task_struct *p = current;
1025 struct mm_struct *mm = p->mm;
1026 struct vm_area_struct *vma;
1027 unsigned long start, end;
1028 unsigned long nr_pte_updates = 0;
1031 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1033 work->next = work; /* protect against double add */
1035 * Who cares about NUMA placement when they're dying.
1037 * NOTE: make sure not to dereference p->mm before this check,
1038 * exit_task_work() happens _after_ exit_mm() so we could be called
1039 * without p->mm even though we still had it when we enqueued this
1042 if (p->flags & PF_EXITING)
1045 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1046 mm->numa_next_scan = now +
1047 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1048 mm->numa_next_reset = now +
1049 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1053 * Reset the scan period if enough time has gone by. Objective is that
1054 * scanning will be reduced if pages are properly placed. As tasks
1055 * can enter different phases this needs to be re-examined. Lacking
1056 * proper tracking of reference behaviour, this blunt hammer is used.
1058 migrate = mm->numa_next_reset;
1059 if (time_after(now, migrate)) {
1060 p->numa_scan_period = task_scan_min(p);
1061 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1062 xchg(&mm->numa_next_reset, next_scan);
1066 * Enforce maximal scan/migration frequency..
1068 migrate = mm->numa_next_scan;
1069 if (time_before(now, migrate))
1072 if (p->numa_scan_period == 0) {
1073 p->numa_scan_period_max = task_scan_max(p);
1074 p->numa_scan_period = task_scan_min(p);
1077 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1078 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1082 * Delay this task enough that another task of this mm will likely win
1083 * the next time around.
1085 p->node_stamp += 2 * TICK_NSEC;
1087 start = mm->numa_scan_offset;
1088 pages = sysctl_numa_balancing_scan_size;
1089 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1093 down_read(&mm->mmap_sem);
1094 vma = find_vma(mm, start);
1096 reset_ptenuma_scan(p);
1100 for (; vma; vma = vma->vm_next) {
1101 if (!vma_migratable(vma))
1104 /* Skip small VMAs. They are not likely to be of relevance */
1105 if (vma->vm_end - vma->vm_start < HPAGE_SIZE)
1109 start = max(start, vma->vm_start);
1110 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1111 end = min(end, vma->vm_end);
1112 nr_pte_updates += change_prot_numa(vma, start, end);
1115 * Scan sysctl_numa_balancing_scan_size but ensure that
1116 * at least one PTE is updated so that unused virtual
1117 * address space is quickly skipped.
1120 pages -= (end - start) >> PAGE_SHIFT;
1125 } while (end != vma->vm_end);
1130 * If the whole process was scanned without updates then no NUMA
1131 * hinting faults are being recorded and scan rate should be lower.
1133 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1134 p->numa_scan_period = min(p->numa_scan_period_max,
1135 p->numa_scan_period << 1);
1137 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1138 mm->numa_next_scan = next_scan;
1142 * It is possible to reach the end of the VMA list but the last few
1143 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1144 * would find the !migratable VMA on the next scan but not reset the
1145 * scanner to the start so check it now.
1148 mm->numa_scan_offset = start;
1150 reset_ptenuma_scan(p);
1151 up_read(&mm->mmap_sem);
1155 * Drive the periodic memory faults..
1157 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1159 struct callback_head *work = &curr->numa_work;
1163 * We don't care about NUMA placement if we don't have memory.
1165 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1169 * Using runtime rather than walltime has the dual advantage that
1170 * we (mostly) drive the selection from busy threads and that the
1171 * task needs to have done some actual work before we bother with
1174 now = curr->se.sum_exec_runtime;
1175 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1177 if (now - curr->node_stamp > period) {
1178 if (!curr->node_stamp)
1179 curr->numa_scan_period = task_scan_min(curr);
1180 curr->node_stamp += period;
1182 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1183 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1184 task_work_add(curr, work, true);
1189 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1192 #endif /* CONFIG_NUMA_BALANCING */
1195 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1197 update_load_add(&cfs_rq->load, se->load.weight);
1198 if (!parent_entity(se))
1199 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1201 if (entity_is_task(se))
1202 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1204 cfs_rq->nr_running++;
1208 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1210 update_load_sub(&cfs_rq->load, se->load.weight);
1211 if (!parent_entity(se))
1212 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1213 if (entity_is_task(se))
1214 list_del_init(&se->group_node);
1215 cfs_rq->nr_running--;
1218 #ifdef CONFIG_FAIR_GROUP_SCHED
1220 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1225 * Use this CPU's actual weight instead of the last load_contribution
1226 * to gain a more accurate current total weight. See
1227 * update_cfs_rq_load_contribution().
1229 tg_weight = atomic_long_read(&tg->load_avg);
1230 tg_weight -= cfs_rq->tg_load_contrib;
1231 tg_weight += cfs_rq->load.weight;
1236 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1238 long tg_weight, load, shares;
1240 tg_weight = calc_tg_weight(tg, cfs_rq);
1241 load = cfs_rq->load.weight;
1243 shares = (tg->shares * load);
1245 shares /= tg_weight;
1247 if (shares < MIN_SHARES)
1248 shares = MIN_SHARES;
1249 if (shares > tg->shares)
1250 shares = tg->shares;
1254 # else /* CONFIG_SMP */
1255 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1259 # endif /* CONFIG_SMP */
1260 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1261 unsigned long weight)
1264 /* commit outstanding execution time */
1265 if (cfs_rq->curr == se)
1266 update_curr(cfs_rq);
1267 account_entity_dequeue(cfs_rq, se);
1270 update_load_set(&se->load, weight);
1273 account_entity_enqueue(cfs_rq, se);
1276 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1278 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1280 struct task_group *tg;
1281 struct sched_entity *se;
1285 se = tg->se[cpu_of(rq_of(cfs_rq))];
1286 if (!se || throttled_hierarchy(cfs_rq))
1289 if (likely(se->load.weight == tg->shares))
1292 shares = calc_cfs_shares(cfs_rq, tg);
1294 reweight_entity(cfs_rq_of(se), se, shares);
1296 #else /* CONFIG_FAIR_GROUP_SCHED */
1297 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1300 #endif /* CONFIG_FAIR_GROUP_SCHED */
1304 * We choose a half-life close to 1 scheduling period.
1305 * Note: The tables below are dependent on this value.
1307 #define LOAD_AVG_PERIOD 32
1308 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1309 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1311 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1312 static const u32 runnable_avg_yN_inv[] = {
1313 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1314 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1315 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1316 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1317 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1318 0x85aac367, 0x82cd8698,
1322 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1323 * over-estimates when re-combining.
1325 static const u32 runnable_avg_yN_sum[] = {
1326 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1327 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1328 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1333 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1335 static __always_inline u64 decay_load(u64 val, u64 n)
1337 unsigned int local_n;
1341 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1344 /* after bounds checking we can collapse to 32-bit */
1348 * As y^PERIOD = 1/2, we can combine
1349 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1350 * With a look-up table which covers k^n (n<PERIOD)
1352 * To achieve constant time decay_load.
1354 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1355 val >>= local_n / LOAD_AVG_PERIOD;
1356 local_n %= LOAD_AVG_PERIOD;
1359 val *= runnable_avg_yN_inv[local_n];
1360 /* We don't use SRR here since we always want to round down. */
1365 * For updates fully spanning n periods, the contribution to runnable
1366 * average will be: \Sum 1024*y^n
1368 * We can compute this reasonably efficiently by combining:
1369 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1371 static u32 __compute_runnable_contrib(u64 n)
1375 if (likely(n <= LOAD_AVG_PERIOD))
1376 return runnable_avg_yN_sum[n];
1377 else if (unlikely(n >= LOAD_AVG_MAX_N))
1378 return LOAD_AVG_MAX;
1380 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1382 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1383 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1385 n -= LOAD_AVG_PERIOD;
1386 } while (n > LOAD_AVG_PERIOD);
1388 contrib = decay_load(contrib, n);
1389 return contrib + runnable_avg_yN_sum[n];
1393 * We can represent the historical contribution to runnable average as the
1394 * coefficients of a geometric series. To do this we sub-divide our runnable
1395 * history into segments of approximately 1ms (1024us); label the segment that
1396 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1398 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1400 * (now) (~1ms ago) (~2ms ago)
1402 * Let u_i denote the fraction of p_i that the entity was runnable.
1404 * We then designate the fractions u_i as our co-efficients, yielding the
1405 * following representation of historical load:
1406 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1408 * We choose y based on the with of a reasonably scheduling period, fixing:
1411 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1412 * approximately half as much as the contribution to load within the last ms
1415 * When a period "rolls over" and we have new u_0`, multiplying the previous
1416 * sum again by y is sufficient to update:
1417 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1418 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1420 static __always_inline int __update_entity_runnable_avg(u64 now,
1421 struct sched_avg *sa,
1425 u32 runnable_contrib;
1426 int delta_w, decayed = 0;
1428 delta = now - sa->last_runnable_update;
1430 * This should only happen when time goes backwards, which it
1431 * unfortunately does during sched clock init when we swap over to TSC.
1433 if ((s64)delta < 0) {
1434 sa->last_runnable_update = now;
1439 * Use 1024ns as the unit of measurement since it's a reasonable
1440 * approximation of 1us and fast to compute.
1445 sa->last_runnable_update = now;
1447 /* delta_w is the amount already accumulated against our next period */
1448 delta_w = sa->runnable_avg_period % 1024;
1449 if (delta + delta_w >= 1024) {
1450 /* period roll-over */
1454 * Now that we know we're crossing a period boundary, figure
1455 * out how much from delta we need to complete the current
1456 * period and accrue it.
1458 delta_w = 1024 - delta_w;
1460 sa->runnable_avg_sum += delta_w;
1461 sa->runnable_avg_period += delta_w;
1465 /* Figure out how many additional periods this update spans */
1466 periods = delta / 1024;
1469 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1471 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1474 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1475 runnable_contrib = __compute_runnable_contrib(periods);
1477 sa->runnable_avg_sum += runnable_contrib;
1478 sa->runnable_avg_period += runnable_contrib;
1481 /* Remainder of delta accrued against u_0` */
1483 sa->runnable_avg_sum += delta;
1484 sa->runnable_avg_period += delta;
1489 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1490 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1492 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1493 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1495 decays -= se->avg.decay_count;
1499 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1500 se->avg.decay_count = 0;
1505 #ifdef CONFIG_FAIR_GROUP_SCHED
1506 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1509 struct task_group *tg = cfs_rq->tg;
1512 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1513 tg_contrib -= cfs_rq->tg_load_contrib;
1515 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1516 atomic_long_add(tg_contrib, &tg->load_avg);
1517 cfs_rq->tg_load_contrib += tg_contrib;
1522 * Aggregate cfs_rq runnable averages into an equivalent task_group
1523 * representation for computing load contributions.
1525 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1526 struct cfs_rq *cfs_rq)
1528 struct task_group *tg = cfs_rq->tg;
1531 /* The fraction of a cpu used by this cfs_rq */
1532 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1533 sa->runnable_avg_period + 1);
1534 contrib -= cfs_rq->tg_runnable_contrib;
1536 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1537 atomic_add(contrib, &tg->runnable_avg);
1538 cfs_rq->tg_runnable_contrib += contrib;
1542 static inline void __update_group_entity_contrib(struct sched_entity *se)
1544 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1545 struct task_group *tg = cfs_rq->tg;
1550 contrib = cfs_rq->tg_load_contrib * tg->shares;
1551 se->avg.load_avg_contrib = div_u64(contrib,
1552 atomic_long_read(&tg->load_avg) + 1);
1555 * For group entities we need to compute a correction term in the case
1556 * that they are consuming <1 cpu so that we would contribute the same
1557 * load as a task of equal weight.
1559 * Explicitly co-ordinating this measurement would be expensive, but
1560 * fortunately the sum of each cpus contribution forms a usable
1561 * lower-bound on the true value.
1563 * Consider the aggregate of 2 contributions. Either they are disjoint
1564 * (and the sum represents true value) or they are disjoint and we are
1565 * understating by the aggregate of their overlap.
1567 * Extending this to N cpus, for a given overlap, the maximum amount we
1568 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1569 * cpus that overlap for this interval and w_i is the interval width.
1571 * On a small machine; the first term is well-bounded which bounds the
1572 * total error since w_i is a subset of the period. Whereas on a
1573 * larger machine, while this first term can be larger, if w_i is the
1574 * of consequential size guaranteed to see n_i*w_i quickly converge to
1575 * our upper bound of 1-cpu.
1577 runnable_avg = atomic_read(&tg->runnable_avg);
1578 if (runnable_avg < NICE_0_LOAD) {
1579 se->avg.load_avg_contrib *= runnable_avg;
1580 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1584 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1585 int force_update) {}
1586 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1587 struct cfs_rq *cfs_rq) {}
1588 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1591 static inline void __update_task_entity_contrib(struct sched_entity *se)
1595 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1596 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1597 contrib /= (se->avg.runnable_avg_period + 1);
1598 se->avg.load_avg_contrib = scale_load(contrib);
1601 /* Compute the current contribution to load_avg by se, return any delta */
1602 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1604 long old_contrib = se->avg.load_avg_contrib;
1606 if (entity_is_task(se)) {
1607 __update_task_entity_contrib(se);
1609 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1610 __update_group_entity_contrib(se);
1613 return se->avg.load_avg_contrib - old_contrib;
1616 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1619 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1620 cfs_rq->blocked_load_avg -= load_contrib;
1622 cfs_rq->blocked_load_avg = 0;
1625 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1627 /* Update a sched_entity's runnable average */
1628 static inline void update_entity_load_avg(struct sched_entity *se,
1631 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1636 * For a group entity we need to use their owned cfs_rq_clock_task() in
1637 * case they are the parent of a throttled hierarchy.
1639 if (entity_is_task(se))
1640 now = cfs_rq_clock_task(cfs_rq);
1642 now = cfs_rq_clock_task(group_cfs_rq(se));
1644 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1647 contrib_delta = __update_entity_load_avg_contrib(se);
1653 cfs_rq->runnable_load_avg += contrib_delta;
1655 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1659 * Decay the load contributed by all blocked children and account this so that
1660 * their contribution may appropriately discounted when they wake up.
1662 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1664 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1667 decays = now - cfs_rq->last_decay;
1668 if (!decays && !force_update)
1671 if (atomic_long_read(&cfs_rq->removed_load)) {
1672 unsigned long removed_load;
1673 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
1674 subtract_blocked_load_contrib(cfs_rq, removed_load);
1678 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1680 atomic64_add(decays, &cfs_rq->decay_counter);
1681 cfs_rq->last_decay = now;
1684 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1687 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1689 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
1690 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1693 /* Add the load generated by se into cfs_rq's child load-average */
1694 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1695 struct sched_entity *se,
1699 * We track migrations using entity decay_count <= 0, on a wake-up
1700 * migration we use a negative decay count to track the remote decays
1701 * accumulated while sleeping.
1703 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
1704 * are seen by enqueue_entity_load_avg() as a migration with an already
1705 * constructed load_avg_contrib.
1707 if (unlikely(se->avg.decay_count <= 0)) {
1708 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
1709 if (se->avg.decay_count) {
1711 * In a wake-up migration we have to approximate the
1712 * time sleeping. This is because we can't synchronize
1713 * clock_task between the two cpus, and it is not
1714 * guaranteed to be read-safe. Instead, we can
1715 * approximate this using our carried decays, which are
1716 * explicitly atomically readable.
1718 se->avg.last_runnable_update -= (-se->avg.decay_count)
1720 update_entity_load_avg(se, 0);
1721 /* Indicate that we're now synchronized and on-rq */
1722 se->avg.decay_count = 0;
1727 * Task re-woke on same cpu (or else migrate_task_rq_fair()
1728 * would have made count negative); we must be careful to avoid
1729 * double-accounting blocked time after synchronizing decays.
1731 se->avg.last_runnable_update += __synchronize_entity_decay(se)
1735 /* migrated tasks did not contribute to our blocked load */
1737 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1738 update_entity_load_avg(se, 0);
1741 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1742 /* we force update consideration on load-balancer moves */
1743 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1747 * Remove se's load from this cfs_rq child load-average, if the entity is
1748 * transitioning to a blocked state we track its projected decay using
1751 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1752 struct sched_entity *se,
1755 update_entity_load_avg(se, 1);
1756 /* we force update consideration on load-balancer moves */
1757 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1759 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1761 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1762 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1763 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1767 * Update the rq's load with the elapsed running time before entering
1768 * idle. if the last scheduled task is not a CFS task, idle_enter will
1769 * be the only way to update the runnable statistic.
1771 void idle_enter_fair(struct rq *this_rq)
1773 update_rq_runnable_avg(this_rq, 1);
1777 * Update the rq's load with the elapsed idle time before a task is
1778 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
1779 * be the only way to update the runnable statistic.
1781 void idle_exit_fair(struct rq *this_rq)
1783 update_rq_runnable_avg(this_rq, 0);
1787 static inline void update_entity_load_avg(struct sched_entity *se,
1788 int update_cfs_rq) {}
1789 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1790 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1791 struct sched_entity *se,
1793 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1794 struct sched_entity *se,
1796 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1797 int force_update) {}
1800 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1802 #ifdef CONFIG_SCHEDSTATS
1803 struct task_struct *tsk = NULL;
1805 if (entity_is_task(se))
1808 if (se->statistics.sleep_start) {
1809 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
1814 if (unlikely(delta > se->statistics.sleep_max))
1815 se->statistics.sleep_max = delta;
1817 se->statistics.sleep_start = 0;
1818 se->statistics.sum_sleep_runtime += delta;
1821 account_scheduler_latency(tsk, delta >> 10, 1);
1822 trace_sched_stat_sleep(tsk, delta);
1825 if (se->statistics.block_start) {
1826 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
1831 if (unlikely(delta > se->statistics.block_max))
1832 se->statistics.block_max = delta;
1834 se->statistics.block_start = 0;
1835 se->statistics.sum_sleep_runtime += delta;
1838 if (tsk->in_iowait) {
1839 se->statistics.iowait_sum += delta;
1840 se->statistics.iowait_count++;
1841 trace_sched_stat_iowait(tsk, delta);
1844 trace_sched_stat_blocked(tsk, delta);
1847 * Blocking time is in units of nanosecs, so shift by
1848 * 20 to get a milliseconds-range estimation of the
1849 * amount of time that the task spent sleeping:
1851 if (unlikely(prof_on == SLEEP_PROFILING)) {
1852 profile_hits(SLEEP_PROFILING,
1853 (void *)get_wchan(tsk),
1856 account_scheduler_latency(tsk, delta >> 10, 0);
1862 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1864 #ifdef CONFIG_SCHED_DEBUG
1865 s64 d = se->vruntime - cfs_rq->min_vruntime;
1870 if (d > 3*sysctl_sched_latency)
1871 schedstat_inc(cfs_rq, nr_spread_over);
1876 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1878 u64 vruntime = cfs_rq->min_vruntime;
1881 * The 'current' period is already promised to the current tasks,
1882 * however the extra weight of the new task will slow them down a
1883 * little, place the new task so that it fits in the slot that
1884 * stays open at the end.
1886 if (initial && sched_feat(START_DEBIT))
1887 vruntime += sched_vslice(cfs_rq, se);
1889 /* sleeps up to a single latency don't count. */
1891 unsigned long thresh = sysctl_sched_latency;
1894 * Halve their sleep time's effect, to allow
1895 * for a gentler effect of sleepers:
1897 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1903 /* ensure we never gain time by being placed backwards. */
1904 se->vruntime = max_vruntime(se->vruntime, vruntime);
1907 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1910 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1913 * Update the normalized vruntime before updating min_vruntime
1914 * through calling update_curr().
1916 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1917 se->vruntime += cfs_rq->min_vruntime;
1920 * Update run-time statistics of the 'current'.
1922 update_curr(cfs_rq);
1923 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1924 account_entity_enqueue(cfs_rq, se);
1925 update_cfs_shares(cfs_rq);
1927 if (flags & ENQUEUE_WAKEUP) {
1928 place_entity(cfs_rq, se, 0);
1929 enqueue_sleeper(cfs_rq, se);
1932 update_stats_enqueue(cfs_rq, se);
1933 check_spread(cfs_rq, se);
1934 if (se != cfs_rq->curr)
1935 __enqueue_entity(cfs_rq, se);
1938 if (cfs_rq->nr_running == 1) {
1939 list_add_leaf_cfs_rq(cfs_rq);
1940 check_enqueue_throttle(cfs_rq);
1944 static void __clear_buddies_last(struct sched_entity *se)
1946 for_each_sched_entity(se) {
1947 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1948 if (cfs_rq->last == se)
1949 cfs_rq->last = NULL;
1955 static void __clear_buddies_next(struct sched_entity *se)
1957 for_each_sched_entity(se) {
1958 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1959 if (cfs_rq->next == se)
1960 cfs_rq->next = NULL;
1966 static void __clear_buddies_skip(struct sched_entity *se)
1968 for_each_sched_entity(se) {
1969 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1970 if (cfs_rq->skip == se)
1971 cfs_rq->skip = NULL;
1977 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1979 if (cfs_rq->last == se)
1980 __clear_buddies_last(se);
1982 if (cfs_rq->next == se)
1983 __clear_buddies_next(se);
1985 if (cfs_rq->skip == se)
1986 __clear_buddies_skip(se);
1989 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1992 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1995 * Update run-time statistics of the 'current'.
1997 update_curr(cfs_rq);
1998 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2000 update_stats_dequeue(cfs_rq, se);
2001 if (flags & DEQUEUE_SLEEP) {
2002 #ifdef CONFIG_SCHEDSTATS
2003 if (entity_is_task(se)) {
2004 struct task_struct *tsk = task_of(se);
2006 if (tsk->state & TASK_INTERRUPTIBLE)
2007 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2008 if (tsk->state & TASK_UNINTERRUPTIBLE)
2009 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2014 clear_buddies(cfs_rq, se);
2016 if (se != cfs_rq->curr)
2017 __dequeue_entity(cfs_rq, se);
2019 account_entity_dequeue(cfs_rq, se);
2022 * Normalize the entity after updating the min_vruntime because the
2023 * update can refer to the ->curr item and we need to reflect this
2024 * movement in our normalized position.
2026 if (!(flags & DEQUEUE_SLEEP))
2027 se->vruntime -= cfs_rq->min_vruntime;
2029 /* return excess runtime on last dequeue */
2030 return_cfs_rq_runtime(cfs_rq);
2032 update_min_vruntime(cfs_rq);
2033 update_cfs_shares(cfs_rq);
2037 * Preempt the current task with a newly woken task if needed:
2040 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2042 unsigned long ideal_runtime, delta_exec;
2043 struct sched_entity *se;
2046 ideal_runtime = sched_slice(cfs_rq, curr);
2047 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2048 if (delta_exec > ideal_runtime) {
2049 resched_task(rq_of(cfs_rq)->curr);
2051 * The current task ran long enough, ensure it doesn't get
2052 * re-elected due to buddy favours.
2054 clear_buddies(cfs_rq, curr);
2059 * Ensure that a task that missed wakeup preemption by a
2060 * narrow margin doesn't have to wait for a full slice.
2061 * This also mitigates buddy induced latencies under load.
2063 if (delta_exec < sysctl_sched_min_granularity)
2066 se = __pick_first_entity(cfs_rq);
2067 delta = curr->vruntime - se->vruntime;
2072 if (delta > ideal_runtime)
2073 resched_task(rq_of(cfs_rq)->curr);
2077 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2079 /* 'current' is not kept within the tree. */
2082 * Any task has to be enqueued before it get to execute on
2083 * a CPU. So account for the time it spent waiting on the
2086 update_stats_wait_end(cfs_rq, se);
2087 __dequeue_entity(cfs_rq, se);
2090 update_stats_curr_start(cfs_rq, se);
2092 #ifdef CONFIG_SCHEDSTATS
2094 * Track our maximum slice length, if the CPU's load is at
2095 * least twice that of our own weight (i.e. dont track it
2096 * when there are only lesser-weight tasks around):
2098 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2099 se->statistics.slice_max = max(se->statistics.slice_max,
2100 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2103 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2107 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2110 * Pick the next process, keeping these things in mind, in this order:
2111 * 1) keep things fair between processes/task groups
2112 * 2) pick the "next" process, since someone really wants that to run
2113 * 3) pick the "last" process, for cache locality
2114 * 4) do not run the "skip" process, if something else is available
2116 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2118 struct sched_entity *se = __pick_first_entity(cfs_rq);
2119 struct sched_entity *left = se;
2122 * Avoid running the skip buddy, if running something else can
2123 * be done without getting too unfair.
2125 if (cfs_rq->skip == se) {
2126 struct sched_entity *second = __pick_next_entity(se);
2127 if (second && wakeup_preempt_entity(second, left) < 1)
2132 * Prefer last buddy, try to return the CPU to a preempted task.
2134 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2138 * Someone really wants this to run. If it's not unfair, run it.
2140 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2143 clear_buddies(cfs_rq, se);
2148 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2150 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2153 * If still on the runqueue then deactivate_task()
2154 * was not called and update_curr() has to be done:
2157 update_curr(cfs_rq);
2159 /* throttle cfs_rqs exceeding runtime */
2160 check_cfs_rq_runtime(cfs_rq);
2162 check_spread(cfs_rq, prev);
2164 update_stats_wait_start(cfs_rq, prev);
2165 /* Put 'current' back into the tree. */
2166 __enqueue_entity(cfs_rq, prev);
2167 /* in !on_rq case, update occurred at dequeue */
2168 update_entity_load_avg(prev, 1);
2170 cfs_rq->curr = NULL;
2174 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2177 * Update run-time statistics of the 'current'.
2179 update_curr(cfs_rq);
2182 * Ensure that runnable average is periodically updated.
2184 update_entity_load_avg(curr, 1);
2185 update_cfs_rq_blocked_load(cfs_rq, 1);
2186 update_cfs_shares(cfs_rq);
2188 #ifdef CONFIG_SCHED_HRTICK
2190 * queued ticks are scheduled to match the slice, so don't bother
2191 * validating it and just reschedule.
2194 resched_task(rq_of(cfs_rq)->curr);
2198 * don't let the period tick interfere with the hrtick preemption
2200 if (!sched_feat(DOUBLE_TICK) &&
2201 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2205 if (cfs_rq->nr_running > 1)
2206 check_preempt_tick(cfs_rq, curr);
2210 /**************************************************
2211 * CFS bandwidth control machinery
2214 #ifdef CONFIG_CFS_BANDWIDTH
2216 #ifdef HAVE_JUMP_LABEL
2217 static struct static_key __cfs_bandwidth_used;
2219 static inline bool cfs_bandwidth_used(void)
2221 return static_key_false(&__cfs_bandwidth_used);
2224 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2226 /* only need to count groups transitioning between enabled/!enabled */
2227 if (enabled && !was_enabled)
2228 static_key_slow_inc(&__cfs_bandwidth_used);
2229 else if (!enabled && was_enabled)
2230 static_key_slow_dec(&__cfs_bandwidth_used);
2232 #else /* HAVE_JUMP_LABEL */
2233 static bool cfs_bandwidth_used(void)
2238 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2239 #endif /* HAVE_JUMP_LABEL */
2242 * default period for cfs group bandwidth.
2243 * default: 0.1s, units: nanoseconds
2245 static inline u64 default_cfs_period(void)
2247 return 100000000ULL;
2250 static inline u64 sched_cfs_bandwidth_slice(void)
2252 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2256 * Replenish runtime according to assigned quota and update expiration time.
2257 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2258 * additional synchronization around rq->lock.
2260 * requires cfs_b->lock
2262 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2266 if (cfs_b->quota == RUNTIME_INF)
2269 now = sched_clock_cpu(smp_processor_id());
2270 cfs_b->runtime = cfs_b->quota;
2271 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2274 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2276 return &tg->cfs_bandwidth;
2279 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2280 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2282 if (unlikely(cfs_rq->throttle_count))
2283 return cfs_rq->throttled_clock_task;
2285 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2288 /* returns 0 on failure to allocate runtime */
2289 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2291 struct task_group *tg = cfs_rq->tg;
2292 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2293 u64 amount = 0, min_amount, expires;
2295 /* note: this is a positive sum as runtime_remaining <= 0 */
2296 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2298 raw_spin_lock(&cfs_b->lock);
2299 if (cfs_b->quota == RUNTIME_INF)
2300 amount = min_amount;
2303 * If the bandwidth pool has become inactive, then at least one
2304 * period must have elapsed since the last consumption.
2305 * Refresh the global state and ensure bandwidth timer becomes
2308 if (!cfs_b->timer_active) {
2309 __refill_cfs_bandwidth_runtime(cfs_b);
2310 __start_cfs_bandwidth(cfs_b);
2313 if (cfs_b->runtime > 0) {
2314 amount = min(cfs_b->runtime, min_amount);
2315 cfs_b->runtime -= amount;
2319 expires = cfs_b->runtime_expires;
2320 raw_spin_unlock(&cfs_b->lock);
2322 cfs_rq->runtime_remaining += amount;
2324 * we may have advanced our local expiration to account for allowed
2325 * spread between our sched_clock and the one on which runtime was
2328 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2329 cfs_rq->runtime_expires = expires;
2331 return cfs_rq->runtime_remaining > 0;
2335 * Note: This depends on the synchronization provided by sched_clock and the
2336 * fact that rq->clock snapshots this value.
2338 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2340 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2342 /* if the deadline is ahead of our clock, nothing to do */
2343 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2346 if (cfs_rq->runtime_remaining < 0)
2350 * If the local deadline has passed we have to consider the
2351 * possibility that our sched_clock is 'fast' and the global deadline
2352 * has not truly expired.
2354 * Fortunately we can check determine whether this the case by checking
2355 * whether the global deadline has advanced.
2358 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2359 /* extend local deadline, drift is bounded above by 2 ticks */
2360 cfs_rq->runtime_expires += TICK_NSEC;
2362 /* global deadline is ahead, expiration has passed */
2363 cfs_rq->runtime_remaining = 0;
2367 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2368 unsigned long delta_exec)
2370 /* dock delta_exec before expiring quota (as it could span periods) */
2371 cfs_rq->runtime_remaining -= delta_exec;
2372 expire_cfs_rq_runtime(cfs_rq);
2374 if (likely(cfs_rq->runtime_remaining > 0))
2378 * if we're unable to extend our runtime we resched so that the active
2379 * hierarchy can be throttled
2381 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2382 resched_task(rq_of(cfs_rq)->curr);
2385 static __always_inline
2386 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2388 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2391 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2394 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2396 return cfs_bandwidth_used() && cfs_rq->throttled;
2399 /* check whether cfs_rq, or any parent, is throttled */
2400 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2402 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2406 * Ensure that neither of the group entities corresponding to src_cpu or
2407 * dest_cpu are members of a throttled hierarchy when performing group
2408 * load-balance operations.
2410 static inline int throttled_lb_pair(struct task_group *tg,
2411 int src_cpu, int dest_cpu)
2413 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2415 src_cfs_rq = tg->cfs_rq[src_cpu];
2416 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2418 return throttled_hierarchy(src_cfs_rq) ||
2419 throttled_hierarchy(dest_cfs_rq);
2422 /* updated child weight may affect parent so we have to do this bottom up */
2423 static int tg_unthrottle_up(struct task_group *tg, void *data)
2425 struct rq *rq = data;
2426 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2428 cfs_rq->throttle_count--;
2430 if (!cfs_rq->throttle_count) {
2431 /* adjust cfs_rq_clock_task() */
2432 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2433 cfs_rq->throttled_clock_task;
2440 static int tg_throttle_down(struct task_group *tg, void *data)
2442 struct rq *rq = data;
2443 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2445 /* group is entering throttled state, stop time */
2446 if (!cfs_rq->throttle_count)
2447 cfs_rq->throttled_clock_task = rq_clock_task(rq);
2448 cfs_rq->throttle_count++;
2453 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2455 struct rq *rq = rq_of(cfs_rq);
2456 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2457 struct sched_entity *se;
2458 long task_delta, dequeue = 1;
2460 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2462 /* freeze hierarchy runnable averages while throttled */
2464 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2467 task_delta = cfs_rq->h_nr_running;
2468 for_each_sched_entity(se) {
2469 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2470 /* throttled entity or throttle-on-deactivate */
2475 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2476 qcfs_rq->h_nr_running -= task_delta;
2478 if (qcfs_rq->load.weight)
2483 rq->nr_running -= task_delta;
2485 cfs_rq->throttled = 1;
2486 cfs_rq->throttled_clock = rq_clock(rq);
2487 raw_spin_lock(&cfs_b->lock);
2488 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2489 raw_spin_unlock(&cfs_b->lock);
2492 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2494 struct rq *rq = rq_of(cfs_rq);
2495 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2496 struct sched_entity *se;
2500 se = cfs_rq->tg->se[cpu_of(rq)];
2502 cfs_rq->throttled = 0;
2504 update_rq_clock(rq);
2506 raw_spin_lock(&cfs_b->lock);
2507 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
2508 list_del_rcu(&cfs_rq->throttled_list);
2509 raw_spin_unlock(&cfs_b->lock);
2511 /* update hierarchical throttle state */
2512 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2514 if (!cfs_rq->load.weight)
2517 task_delta = cfs_rq->h_nr_running;
2518 for_each_sched_entity(se) {
2522 cfs_rq = cfs_rq_of(se);
2524 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2525 cfs_rq->h_nr_running += task_delta;
2527 if (cfs_rq_throttled(cfs_rq))
2532 rq->nr_running += task_delta;
2534 /* determine whether we need to wake up potentially idle cpu */
2535 if (rq->curr == rq->idle && rq->cfs.nr_running)
2536 resched_task(rq->curr);
2539 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2540 u64 remaining, u64 expires)
2542 struct cfs_rq *cfs_rq;
2543 u64 runtime = remaining;
2546 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2548 struct rq *rq = rq_of(cfs_rq);
2550 raw_spin_lock(&rq->lock);
2551 if (!cfs_rq_throttled(cfs_rq))
2554 runtime = -cfs_rq->runtime_remaining + 1;
2555 if (runtime > remaining)
2556 runtime = remaining;
2557 remaining -= runtime;
2559 cfs_rq->runtime_remaining += runtime;
2560 cfs_rq->runtime_expires = expires;
2562 /* we check whether we're throttled above */
2563 if (cfs_rq->runtime_remaining > 0)
2564 unthrottle_cfs_rq(cfs_rq);
2567 raw_spin_unlock(&rq->lock);
2578 * Responsible for refilling a task_group's bandwidth and unthrottling its
2579 * cfs_rqs as appropriate. If there has been no activity within the last
2580 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2581 * used to track this state.
2583 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2585 u64 runtime, runtime_expires;
2586 int idle = 1, throttled;
2588 raw_spin_lock(&cfs_b->lock);
2589 /* no need to continue the timer with no bandwidth constraint */
2590 if (cfs_b->quota == RUNTIME_INF)
2593 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2594 /* idle depends on !throttled (for the case of a large deficit) */
2595 idle = cfs_b->idle && !throttled;
2596 cfs_b->nr_periods += overrun;
2598 /* if we're going inactive then everything else can be deferred */
2602 __refill_cfs_bandwidth_runtime(cfs_b);
2605 /* mark as potentially idle for the upcoming period */
2610 /* account preceding periods in which throttling occurred */
2611 cfs_b->nr_throttled += overrun;
2614 * There are throttled entities so we must first use the new bandwidth
2615 * to unthrottle them before making it generally available. This
2616 * ensures that all existing debts will be paid before a new cfs_rq is
2619 runtime = cfs_b->runtime;
2620 runtime_expires = cfs_b->runtime_expires;
2624 * This check is repeated as we are holding onto the new bandwidth
2625 * while we unthrottle. This can potentially race with an unthrottled
2626 * group trying to acquire new bandwidth from the global pool.
2628 while (throttled && runtime > 0) {
2629 raw_spin_unlock(&cfs_b->lock);
2630 /* we can't nest cfs_b->lock while distributing bandwidth */
2631 runtime = distribute_cfs_runtime(cfs_b, runtime,
2633 raw_spin_lock(&cfs_b->lock);
2635 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2638 /* return (any) remaining runtime */
2639 cfs_b->runtime = runtime;
2641 * While we are ensured activity in the period following an
2642 * unthrottle, this also covers the case in which the new bandwidth is
2643 * insufficient to cover the existing bandwidth deficit. (Forcing the
2644 * timer to remain active while there are any throttled entities.)
2649 cfs_b->timer_active = 0;
2650 raw_spin_unlock(&cfs_b->lock);
2655 /* a cfs_rq won't donate quota below this amount */
2656 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2657 /* minimum remaining period time to redistribute slack quota */
2658 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2659 /* how long we wait to gather additional slack before distributing */
2660 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2662 /* are we near the end of the current quota period? */
2663 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2665 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2668 /* if the call-back is running a quota refresh is already occurring */
2669 if (hrtimer_callback_running(refresh_timer))
2672 /* is a quota refresh about to occur? */
2673 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2674 if (remaining < min_expire)
2680 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2682 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2684 /* if there's a quota refresh soon don't bother with slack */
2685 if (runtime_refresh_within(cfs_b, min_left))
2688 start_bandwidth_timer(&cfs_b->slack_timer,
2689 ns_to_ktime(cfs_bandwidth_slack_period));
2692 /* we know any runtime found here is valid as update_curr() precedes return */
2693 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2695 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2696 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2698 if (slack_runtime <= 0)
2701 raw_spin_lock(&cfs_b->lock);
2702 if (cfs_b->quota != RUNTIME_INF &&
2703 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2704 cfs_b->runtime += slack_runtime;
2706 /* we are under rq->lock, defer unthrottling using a timer */
2707 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2708 !list_empty(&cfs_b->throttled_cfs_rq))
2709 start_cfs_slack_bandwidth(cfs_b);
2711 raw_spin_unlock(&cfs_b->lock);
2713 /* even if it's not valid for return we don't want to try again */
2714 cfs_rq->runtime_remaining -= slack_runtime;
2717 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2719 if (!cfs_bandwidth_used())
2722 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2725 __return_cfs_rq_runtime(cfs_rq);
2729 * This is done with a timer (instead of inline with bandwidth return) since
2730 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2732 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2734 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2737 /* confirm we're still not at a refresh boundary */
2738 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2741 raw_spin_lock(&cfs_b->lock);
2742 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2743 runtime = cfs_b->runtime;
2746 expires = cfs_b->runtime_expires;
2747 raw_spin_unlock(&cfs_b->lock);
2752 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2754 raw_spin_lock(&cfs_b->lock);
2755 if (expires == cfs_b->runtime_expires)
2756 cfs_b->runtime = runtime;
2757 raw_spin_unlock(&cfs_b->lock);
2761 * When a group wakes up we want to make sure that its quota is not already
2762 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2763 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2765 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2767 if (!cfs_bandwidth_used())
2770 /* an active group must be handled by the update_curr()->put() path */
2771 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2774 /* ensure the group is not already throttled */
2775 if (cfs_rq_throttled(cfs_rq))
2778 /* update runtime allocation */
2779 account_cfs_rq_runtime(cfs_rq, 0);
2780 if (cfs_rq->runtime_remaining <= 0)
2781 throttle_cfs_rq(cfs_rq);
2784 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2785 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2787 if (!cfs_bandwidth_used())
2790 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2794 * it's possible for a throttled entity to be forced into a running
2795 * state (e.g. set_curr_task), in this case we're finished.
2797 if (cfs_rq_throttled(cfs_rq))
2800 throttle_cfs_rq(cfs_rq);
2803 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2805 struct cfs_bandwidth *cfs_b =
2806 container_of(timer, struct cfs_bandwidth, slack_timer);
2807 do_sched_cfs_slack_timer(cfs_b);
2809 return HRTIMER_NORESTART;
2812 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2814 struct cfs_bandwidth *cfs_b =
2815 container_of(timer, struct cfs_bandwidth, period_timer);
2821 now = hrtimer_cb_get_time(timer);
2822 overrun = hrtimer_forward(timer, now, cfs_b->period);
2827 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2830 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2833 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2835 raw_spin_lock_init(&cfs_b->lock);
2837 cfs_b->quota = RUNTIME_INF;
2838 cfs_b->period = ns_to_ktime(default_cfs_period());
2840 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2841 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2842 cfs_b->period_timer.function = sched_cfs_period_timer;
2843 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2844 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2847 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2849 cfs_rq->runtime_enabled = 0;
2850 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2853 /* requires cfs_b->lock, may release to reprogram timer */
2854 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2857 * The timer may be active because we're trying to set a new bandwidth
2858 * period or because we're racing with the tear-down path
2859 * (timer_active==0 becomes visible before the hrtimer call-back
2860 * terminates). In either case we ensure that it's re-programmed
2862 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2863 raw_spin_unlock(&cfs_b->lock);
2864 /* ensure cfs_b->lock is available while we wait */
2865 hrtimer_cancel(&cfs_b->period_timer);
2867 raw_spin_lock(&cfs_b->lock);
2868 /* if someone else restarted the timer then we're done */
2869 if (cfs_b->timer_active)
2873 cfs_b->timer_active = 1;
2874 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2877 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2879 hrtimer_cancel(&cfs_b->period_timer);
2880 hrtimer_cancel(&cfs_b->slack_timer);
2883 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
2885 struct cfs_rq *cfs_rq;
2887 for_each_leaf_cfs_rq(rq, cfs_rq) {
2888 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2890 if (!cfs_rq->runtime_enabled)
2894 * clock_task is not advancing so we just need to make sure
2895 * there's some valid quota amount
2897 cfs_rq->runtime_remaining = cfs_b->quota;
2898 if (cfs_rq_throttled(cfs_rq))
2899 unthrottle_cfs_rq(cfs_rq);
2903 #else /* CONFIG_CFS_BANDWIDTH */
2904 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2906 return rq_clock_task(rq_of(cfs_rq));
2909 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2910 unsigned long delta_exec) {}
2911 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2912 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2913 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2915 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2920 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2925 static inline int throttled_lb_pair(struct task_group *tg,
2926 int src_cpu, int dest_cpu)
2931 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2933 #ifdef CONFIG_FAIR_GROUP_SCHED
2934 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2937 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2941 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2942 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2944 #endif /* CONFIG_CFS_BANDWIDTH */
2946 /**************************************************
2947 * CFS operations on tasks:
2950 #ifdef CONFIG_SCHED_HRTICK
2951 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2953 struct sched_entity *se = &p->se;
2954 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2956 WARN_ON(task_rq(p) != rq);
2958 if (cfs_rq->nr_running > 1) {
2959 u64 slice = sched_slice(cfs_rq, se);
2960 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2961 s64 delta = slice - ran;
2970 * Don't schedule slices shorter than 10000ns, that just
2971 * doesn't make sense. Rely on vruntime for fairness.
2974 delta = max_t(s64, 10000LL, delta);
2976 hrtick_start(rq, delta);
2981 * called from enqueue/dequeue and updates the hrtick when the
2982 * current task is from our class and nr_running is low enough
2985 static void hrtick_update(struct rq *rq)
2987 struct task_struct *curr = rq->curr;
2989 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2992 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2993 hrtick_start_fair(rq, curr);
2995 #else /* !CONFIG_SCHED_HRTICK */
2997 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3001 static inline void hrtick_update(struct rq *rq)
3007 * The enqueue_task method is called before nr_running is
3008 * increased. Here we update the fair scheduling stats and
3009 * then put the task into the rbtree:
3012 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3014 struct cfs_rq *cfs_rq;
3015 struct sched_entity *se = &p->se;
3017 for_each_sched_entity(se) {
3020 cfs_rq = cfs_rq_of(se);
3021 enqueue_entity(cfs_rq, se, flags);
3024 * end evaluation on encountering a throttled cfs_rq
3026 * note: in the case of encountering a throttled cfs_rq we will
3027 * post the final h_nr_running increment below.
3029 if (cfs_rq_throttled(cfs_rq))
3031 cfs_rq->h_nr_running++;
3033 flags = ENQUEUE_WAKEUP;
3036 for_each_sched_entity(se) {
3037 cfs_rq = cfs_rq_of(se);
3038 cfs_rq->h_nr_running++;
3040 if (cfs_rq_throttled(cfs_rq))
3043 update_cfs_shares(cfs_rq);
3044 update_entity_load_avg(se, 1);
3048 update_rq_runnable_avg(rq, rq->nr_running);
3054 static void set_next_buddy(struct sched_entity *se);
3057 * The dequeue_task method is called before nr_running is
3058 * decreased. We remove the task from the rbtree and
3059 * update the fair scheduling stats:
3061 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3063 struct cfs_rq *cfs_rq;
3064 struct sched_entity *se = &p->se;
3065 int task_sleep = flags & DEQUEUE_SLEEP;
3067 for_each_sched_entity(se) {
3068 cfs_rq = cfs_rq_of(se);
3069 dequeue_entity(cfs_rq, se, flags);
3072 * end evaluation on encountering a throttled cfs_rq
3074 * note: in the case of encountering a throttled cfs_rq we will
3075 * post the final h_nr_running decrement below.
3077 if (cfs_rq_throttled(cfs_rq))
3079 cfs_rq->h_nr_running--;
3081 /* Don't dequeue parent if it has other entities besides us */
3082 if (cfs_rq->load.weight) {
3084 * Bias pick_next to pick a task from this cfs_rq, as
3085 * p is sleeping when it is within its sched_slice.
3087 if (task_sleep && parent_entity(se))
3088 set_next_buddy(parent_entity(se));
3090 /* avoid re-evaluating load for this entity */
3091 se = parent_entity(se);
3094 flags |= DEQUEUE_SLEEP;
3097 for_each_sched_entity(se) {
3098 cfs_rq = cfs_rq_of(se);
3099 cfs_rq->h_nr_running--;
3101 if (cfs_rq_throttled(cfs_rq))
3104 update_cfs_shares(cfs_rq);
3105 update_entity_load_avg(se, 1);
3110 update_rq_runnable_avg(rq, 1);
3116 /* Used instead of source_load when we know the type == 0 */
3117 static unsigned long weighted_cpuload(const int cpu)
3119 return cpu_rq(cpu)->cfs.runnable_load_avg;
3123 * Return a low guess at the load of a migration-source cpu weighted
3124 * according to the scheduling class and "nice" value.
3126 * We want to under-estimate the load of migration sources, to
3127 * balance conservatively.
3129 static unsigned long source_load(int cpu, int type)
3131 struct rq *rq = cpu_rq(cpu);
3132 unsigned long total = weighted_cpuload(cpu);
3134 if (type == 0 || !sched_feat(LB_BIAS))
3137 return min(rq->cpu_load[type-1], total);
3141 * Return a high guess at the load of a migration-target cpu weighted
3142 * according to the scheduling class and "nice" value.
3144 static unsigned long target_load(int cpu, int type)
3146 struct rq *rq = cpu_rq(cpu);
3147 unsigned long total = weighted_cpuload(cpu);
3149 if (type == 0 || !sched_feat(LB_BIAS))
3152 return max(rq->cpu_load[type-1], total);
3155 static unsigned long power_of(int cpu)
3157 return cpu_rq(cpu)->cpu_power;
3160 static unsigned long cpu_avg_load_per_task(int cpu)
3162 struct rq *rq = cpu_rq(cpu);
3163 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3164 unsigned long load_avg = rq->cfs.runnable_load_avg;
3167 return load_avg / nr_running;
3172 static void record_wakee(struct task_struct *p)
3175 * Rough decay (wiping) for cost saving, don't worry
3176 * about the boundary, really active task won't care
3179 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3180 current->wakee_flips = 0;
3181 current->wakee_flip_decay_ts = jiffies;
3184 if (current->last_wakee != p) {
3185 current->last_wakee = p;
3186 current->wakee_flips++;
3190 static void task_waking_fair(struct task_struct *p)
3192 struct sched_entity *se = &p->se;
3193 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3196 #ifndef CONFIG_64BIT
3197 u64 min_vruntime_copy;
3200 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3202 min_vruntime = cfs_rq->min_vruntime;
3203 } while (min_vruntime != min_vruntime_copy);
3205 min_vruntime = cfs_rq->min_vruntime;
3208 se->vruntime -= min_vruntime;
3212 #ifdef CONFIG_FAIR_GROUP_SCHED
3214 * effective_load() calculates the load change as seen from the root_task_group
3216 * Adding load to a group doesn't make a group heavier, but can cause movement
3217 * of group shares between cpus. Assuming the shares were perfectly aligned one
3218 * can calculate the shift in shares.
3220 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3221 * on this @cpu and results in a total addition (subtraction) of @wg to the
3222 * total group weight.
3224 * Given a runqueue weight distribution (rw_i) we can compute a shares
3225 * distribution (s_i) using:
3227 * s_i = rw_i / \Sum rw_j (1)
3229 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3230 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3231 * shares distribution (s_i):
3233 * rw_i = { 2, 4, 1, 0 }
3234 * s_i = { 2/7, 4/7, 1/7, 0 }
3236 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3237 * task used to run on and the CPU the waker is running on), we need to
3238 * compute the effect of waking a task on either CPU and, in case of a sync
3239 * wakeup, compute the effect of the current task going to sleep.
3241 * So for a change of @wl to the local @cpu with an overall group weight change
3242 * of @wl we can compute the new shares distribution (s'_i) using:
3244 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3246 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3247 * differences in waking a task to CPU 0. The additional task changes the
3248 * weight and shares distributions like:
3250 * rw'_i = { 3, 4, 1, 0 }
3251 * s'_i = { 3/8, 4/8, 1/8, 0 }
3253 * We can then compute the difference in effective weight by using:
3255 * dw_i = S * (s'_i - s_i) (3)
3257 * Where 'S' is the group weight as seen by its parent.
3259 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3260 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3261 * 4/7) times the weight of the group.
3263 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3265 struct sched_entity *se = tg->se[cpu];
3267 if (!tg->parent) /* the trivial, non-cgroup case */
3270 for_each_sched_entity(se) {
3276 * W = @wg + \Sum rw_j
3278 W = wg + calc_tg_weight(tg, se->my_q);
3283 w = se->my_q->load.weight + wl;
3286 * wl = S * s'_i; see (2)
3289 wl = (w * tg->shares) / W;
3294 * Per the above, wl is the new se->load.weight value; since
3295 * those are clipped to [MIN_SHARES, ...) do so now. See
3296 * calc_cfs_shares().
3298 if (wl < MIN_SHARES)
3302 * wl = dw_i = S * (s'_i - s_i); see (3)
3304 wl -= se->load.weight;
3307 * Recursively apply this logic to all parent groups to compute
3308 * the final effective load change on the root group. Since
3309 * only the @tg group gets extra weight, all parent groups can
3310 * only redistribute existing shares. @wl is the shift in shares
3311 * resulting from this level per the above.
3320 static inline unsigned long effective_load(struct task_group *tg, int cpu,
3321 unsigned long wl, unsigned long wg)
3328 static int wake_wide(struct task_struct *p)
3330 int factor = this_cpu_read(sd_llc_size);
3333 * Yeah, it's the switching-frequency, could means many wakee or
3334 * rapidly switch, use factor here will just help to automatically
3335 * adjust the loose-degree, so bigger node will lead to more pull.
3337 if (p->wakee_flips > factor) {
3339 * wakee is somewhat hot, it needs certain amount of cpu
3340 * resource, so if waker is far more hot, prefer to leave
3343 if (current->wakee_flips > (factor * p->wakee_flips))
3350 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3352 s64 this_load, load;
3353 int idx, this_cpu, prev_cpu;
3354 unsigned long tl_per_task;
3355 struct task_group *tg;
3356 unsigned long weight;
3360 * If we wake multiple tasks be careful to not bounce
3361 * ourselves around too much.
3367 this_cpu = smp_processor_id();
3368 prev_cpu = task_cpu(p);
3369 load = source_load(prev_cpu, idx);
3370 this_load = target_load(this_cpu, idx);
3373 * If sync wakeup then subtract the (maximum possible)
3374 * effect of the currently running task from the load
3375 * of the current CPU:
3378 tg = task_group(current);
3379 weight = current->se.load.weight;
3381 this_load += effective_load(tg, this_cpu, -weight, -weight);
3382 load += effective_load(tg, prev_cpu, 0, -weight);
3386 weight = p->se.load.weight;
3389 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3390 * due to the sync cause above having dropped this_load to 0, we'll
3391 * always have an imbalance, but there's really nothing you can do
3392 * about that, so that's good too.
3394 * Otherwise check if either cpus are near enough in load to allow this
3395 * task to be woken on this_cpu.
3397 if (this_load > 0) {
3398 s64 this_eff_load, prev_eff_load;
3400 this_eff_load = 100;
3401 this_eff_load *= power_of(prev_cpu);
3402 this_eff_load *= this_load +
3403 effective_load(tg, this_cpu, weight, weight);
3405 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3406 prev_eff_load *= power_of(this_cpu);
3407 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3409 balanced = this_eff_load <= prev_eff_load;
3414 * If the currently running task will sleep within
3415 * a reasonable amount of time then attract this newly
3418 if (sync && balanced)
3421 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3422 tl_per_task = cpu_avg_load_per_task(this_cpu);
3425 (this_load <= load &&
3426 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3428 * This domain has SD_WAKE_AFFINE and
3429 * p is cache cold in this domain, and
3430 * there is no bad imbalance.
3432 schedstat_inc(sd, ttwu_move_affine);
3433 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3441 * find_idlest_group finds and returns the least busy CPU group within the
3444 static struct sched_group *
3445 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3446 int this_cpu, int load_idx)
3448 struct sched_group *idlest = NULL, *group = sd->groups;
3449 unsigned long min_load = ULONG_MAX, this_load = 0;
3450 int imbalance = 100 + (sd->imbalance_pct-100)/2;
3453 unsigned long load, avg_load;
3457 /* Skip over this group if it has no CPUs allowed */
3458 if (!cpumask_intersects(sched_group_cpus(group),
3459 tsk_cpus_allowed(p)))
3462 local_group = cpumask_test_cpu(this_cpu,
3463 sched_group_cpus(group));
3465 /* Tally up the load of all CPUs in the group */
3468 for_each_cpu(i, sched_group_cpus(group)) {
3469 /* Bias balancing toward cpus of our domain */
3471 load = source_load(i, load_idx);
3473 load = target_load(i, load_idx);
3478 /* Adjust by relative CPU power of the group */
3479 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
3482 this_load = avg_load;
3483 } else if (avg_load < min_load) {
3484 min_load = avg_load;
3487 } while (group = group->next, group != sd->groups);
3489 if (!idlest || 100*this_load < imbalance*min_load)
3495 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3498 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3500 unsigned long load, min_load = ULONG_MAX;
3504 /* Traverse only the allowed CPUs */
3505 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3506 load = weighted_cpuload(i);
3508 if (load < min_load || (load == min_load && i == this_cpu)) {
3518 * Try and locate an idle CPU in the sched_domain.
3520 static int select_idle_sibling(struct task_struct *p, int target)
3522 struct sched_domain *sd;
3523 struct sched_group *sg;
3524 int i = task_cpu(p);
3526 if (idle_cpu(target))
3530 * If the prevous cpu is cache affine and idle, don't be stupid.
3532 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
3536 * Otherwise, iterate the domains and find an elegible idle cpu.
3538 sd = rcu_dereference(per_cpu(sd_llc, target));
3539 for_each_lower_domain(sd) {
3542 if (!cpumask_intersects(sched_group_cpus(sg),
3543 tsk_cpus_allowed(p)))
3546 for_each_cpu(i, sched_group_cpus(sg)) {
3547 if (i == target || !idle_cpu(i))
3551 target = cpumask_first_and(sched_group_cpus(sg),
3552 tsk_cpus_allowed(p));
3556 } while (sg != sd->groups);
3563 * sched_balance_self: balance the current task (running on cpu) in domains
3564 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3567 * Balance, ie. select the least loaded group.
3569 * Returns the target CPU number, or the same CPU if no balancing is needed.
3571 * preempt must be disabled.
3574 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3576 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3577 int cpu = smp_processor_id();
3578 int prev_cpu = task_cpu(p);
3580 int want_affine = 0;
3581 int sync = wake_flags & WF_SYNC;
3583 if (p->nr_cpus_allowed == 1)
3586 if (sd_flag & SD_BALANCE_WAKE) {
3587 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3593 for_each_domain(cpu, tmp) {
3594 if (!(tmp->flags & SD_LOAD_BALANCE))
3598 * If both cpu and prev_cpu are part of this domain,
3599 * cpu is a valid SD_WAKE_AFFINE target.
3601 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3602 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3607 if (tmp->flags & sd_flag)
3612 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3615 new_cpu = select_idle_sibling(p, prev_cpu);
3620 int load_idx = sd->forkexec_idx;
3621 struct sched_group *group;
3624 if (!(sd->flags & sd_flag)) {
3629 if (sd_flag & SD_BALANCE_WAKE)
3630 load_idx = sd->wake_idx;
3632 group = find_idlest_group(sd, p, cpu, load_idx);
3638 new_cpu = find_idlest_cpu(group, p, cpu);
3639 if (new_cpu == -1 || new_cpu == cpu) {
3640 /* Now try balancing at a lower domain level of cpu */
3645 /* Now try balancing at a lower domain level of new_cpu */
3647 weight = sd->span_weight;
3649 for_each_domain(cpu, tmp) {
3650 if (weight <= tmp->span_weight)
3652 if (tmp->flags & sd_flag)
3655 /* while loop will break here if sd == NULL */
3664 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3665 * cfs_rq_of(p) references at time of call are still valid and identify the
3666 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3667 * other assumptions, including the state of rq->lock, should be made.
3670 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3672 struct sched_entity *se = &p->se;
3673 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3676 * Load tracking: accumulate removed load so that it can be processed
3677 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3678 * to blocked load iff they have a positive decay-count. It can never
3679 * be negative here since on-rq tasks have decay-count == 0.
3681 if (se->avg.decay_count) {
3682 se->avg.decay_count = -__synchronize_entity_decay(se);
3683 atomic_long_add(se->avg.load_avg_contrib,
3684 &cfs_rq->removed_load);
3687 #endif /* CONFIG_SMP */
3689 static unsigned long
3690 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3692 unsigned long gran = sysctl_sched_wakeup_granularity;
3695 * Since its curr running now, convert the gran from real-time
3696 * to virtual-time in his units.
3698 * By using 'se' instead of 'curr' we penalize light tasks, so
3699 * they get preempted easier. That is, if 'se' < 'curr' then
3700 * the resulting gran will be larger, therefore penalizing the
3701 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3702 * be smaller, again penalizing the lighter task.
3704 * This is especially important for buddies when the leftmost
3705 * task is higher priority than the buddy.
3707 return calc_delta_fair(gran, se);
3711 * Should 'se' preempt 'curr'.
3725 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3727 s64 gran, vdiff = curr->vruntime - se->vruntime;
3732 gran = wakeup_gran(curr, se);
3739 static void set_last_buddy(struct sched_entity *se)
3741 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3744 for_each_sched_entity(se)
3745 cfs_rq_of(se)->last = se;
3748 static void set_next_buddy(struct sched_entity *se)
3750 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3753 for_each_sched_entity(se)
3754 cfs_rq_of(se)->next = se;
3757 static void set_skip_buddy(struct sched_entity *se)
3759 for_each_sched_entity(se)
3760 cfs_rq_of(se)->skip = se;
3764 * Preempt the current task with a newly woken task if needed:
3766 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3768 struct task_struct *curr = rq->curr;
3769 struct sched_entity *se = &curr->se, *pse = &p->se;
3770 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3771 int scale = cfs_rq->nr_running >= sched_nr_latency;
3772 int next_buddy_marked = 0;
3774 if (unlikely(se == pse))
3778 * This is possible from callers such as move_task(), in which we
3779 * unconditionally check_prempt_curr() after an enqueue (which may have
3780 * lead to a throttle). This both saves work and prevents false
3781 * next-buddy nomination below.
3783 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3786 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3787 set_next_buddy(pse);
3788 next_buddy_marked = 1;
3792 * We can come here with TIF_NEED_RESCHED already set from new task
3795 * Note: this also catches the edge-case of curr being in a throttled
3796 * group (e.g. via set_curr_task), since update_curr() (in the
3797 * enqueue of curr) will have resulted in resched being set. This
3798 * prevents us from potentially nominating it as a false LAST_BUDDY
3801 if (test_tsk_need_resched(curr))
3804 /* Idle tasks are by definition preempted by non-idle tasks. */
3805 if (unlikely(curr->policy == SCHED_IDLE) &&
3806 likely(p->policy != SCHED_IDLE))
3810 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3811 * is driven by the tick):
3813 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3816 find_matching_se(&se, &pse);
3817 update_curr(cfs_rq_of(se));
3819 if (wakeup_preempt_entity(se, pse) == 1) {
3821 * Bias pick_next to pick the sched entity that is
3822 * triggering this preemption.
3824 if (!next_buddy_marked)
3825 set_next_buddy(pse);
3834 * Only set the backward buddy when the current task is still
3835 * on the rq. This can happen when a wakeup gets interleaved
3836 * with schedule on the ->pre_schedule() or idle_balance()
3837 * point, either of which can * drop the rq lock.
3839 * Also, during early boot the idle thread is in the fair class,
3840 * for obvious reasons its a bad idea to schedule back to it.
3842 if (unlikely(!se->on_rq || curr == rq->idle))
3845 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3849 static struct task_struct *pick_next_task_fair(struct rq *rq)
3851 struct task_struct *p;
3852 struct cfs_rq *cfs_rq = &rq->cfs;
3853 struct sched_entity *se;
3855 if (!cfs_rq->nr_running)
3859 se = pick_next_entity(cfs_rq);
3860 set_next_entity(cfs_rq, se);
3861 cfs_rq = group_cfs_rq(se);
3865 if (hrtick_enabled(rq))
3866 hrtick_start_fair(rq, p);
3872 * Account for a descheduled task:
3874 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3876 struct sched_entity *se = &prev->se;
3877 struct cfs_rq *cfs_rq;
3879 for_each_sched_entity(se) {
3880 cfs_rq = cfs_rq_of(se);
3881 put_prev_entity(cfs_rq, se);
3886 * sched_yield() is very simple
3888 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3890 static void yield_task_fair(struct rq *rq)
3892 struct task_struct *curr = rq->curr;
3893 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3894 struct sched_entity *se = &curr->se;
3897 * Are we the only task in the tree?
3899 if (unlikely(rq->nr_running == 1))
3902 clear_buddies(cfs_rq, se);
3904 if (curr->policy != SCHED_BATCH) {
3905 update_rq_clock(rq);
3907 * Update run-time statistics of the 'current'.
3909 update_curr(cfs_rq);
3911 * Tell update_rq_clock() that we've just updated,
3912 * so we don't do microscopic update in schedule()
3913 * and double the fastpath cost.
3915 rq->skip_clock_update = 1;
3921 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3923 struct sched_entity *se = &p->se;
3925 /* throttled hierarchies are not runnable */
3926 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3929 /* Tell the scheduler that we'd really like pse to run next. */
3932 yield_task_fair(rq);
3938 /**************************************************
3939 * Fair scheduling class load-balancing methods.
3943 * The purpose of load-balancing is to achieve the same basic fairness the
3944 * per-cpu scheduler provides, namely provide a proportional amount of compute
3945 * time to each task. This is expressed in the following equation:
3947 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3949 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3950 * W_i,0 is defined as:
3952 * W_i,0 = \Sum_j w_i,j (2)
3954 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3955 * is derived from the nice value as per prio_to_weight[].
3957 * The weight average is an exponential decay average of the instantaneous
3960 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3962 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3963 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3964 * can also include other factors [XXX].
3966 * To achieve this balance we define a measure of imbalance which follows
3967 * directly from (1):
3969 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3971 * We them move tasks around to minimize the imbalance. In the continuous
3972 * function space it is obvious this converges, in the discrete case we get
3973 * a few fun cases generally called infeasible weight scenarios.
3976 * - infeasible weights;
3977 * - local vs global optima in the discrete case. ]
3982 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3983 * for all i,j solution, we create a tree of cpus that follows the hardware
3984 * topology where each level pairs two lower groups (or better). This results
3985 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3986 * tree to only the first of the previous level and we decrease the frequency
3987 * of load-balance at each level inv. proportional to the number of cpus in
3993 * \Sum { --- * --- * 2^i } = O(n) (5)
3995 * `- size of each group
3996 * | | `- number of cpus doing load-balance
3998 * `- sum over all levels
4000 * Coupled with a limit on how many tasks we can migrate every balance pass,
4001 * this makes (5) the runtime complexity of the balancer.
4003 * An important property here is that each CPU is still (indirectly) connected
4004 * to every other cpu in at most O(log n) steps:
4006 * The adjacency matrix of the resulting graph is given by:
4009 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4012 * And you'll find that:
4014 * A^(log_2 n)_i,j != 0 for all i,j (7)
4016 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4017 * The task movement gives a factor of O(m), giving a convergence complexity
4020 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4025 * In order to avoid CPUs going idle while there's still work to do, new idle
4026 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4027 * tree itself instead of relying on other CPUs to bring it work.
4029 * This adds some complexity to both (5) and (8) but it reduces the total idle
4037 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4040 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4045 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4047 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4049 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4052 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4053 * rewrite all of this once again.]
4056 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4058 #define LBF_ALL_PINNED 0x01
4059 #define LBF_NEED_BREAK 0x02
4060 #define LBF_DST_PINNED 0x04
4061 #define LBF_SOME_PINNED 0x08
4064 struct sched_domain *sd;
4072 struct cpumask *dst_grpmask;
4074 enum cpu_idle_type idle;
4076 /* The set of CPUs under consideration for load-balancing */
4077 struct cpumask *cpus;
4082 unsigned int loop_break;
4083 unsigned int loop_max;
4087 * move_task - move a task from one runqueue to another runqueue.
4088 * Both runqueues must be locked.
4090 static void move_task(struct task_struct *p, struct lb_env *env)
4092 deactivate_task(env->src_rq, p, 0);
4093 set_task_cpu(p, env->dst_cpu);
4094 activate_task(env->dst_rq, p, 0);
4095 check_preempt_curr(env->dst_rq, p, 0);
4099 * Is this task likely cache-hot:
4102 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4106 if (p->sched_class != &fair_sched_class)
4109 if (unlikely(p->policy == SCHED_IDLE))
4113 * Buddy candidates are cache hot:
4115 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4116 (&p->se == cfs_rq_of(&p->se)->next ||
4117 &p->se == cfs_rq_of(&p->se)->last))
4120 if (sysctl_sched_migration_cost == -1)
4122 if (sysctl_sched_migration_cost == 0)
4125 delta = now - p->se.exec_start;
4127 return delta < (s64)sysctl_sched_migration_cost;
4130 #ifdef CONFIG_NUMA_BALANCING
4131 /* Returns true if the destination node has incurred more faults */
4132 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4134 int src_nid, dst_nid;
4136 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4137 !(env->sd->flags & SD_NUMA)) {
4141 src_nid = cpu_to_node(env->src_cpu);
4142 dst_nid = cpu_to_node(env->dst_cpu);
4144 if (src_nid == dst_nid ||
4145 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4148 if (dst_nid == p->numa_preferred_nid ||
4149 p->numa_faults[dst_nid] > p->numa_faults[src_nid])
4156 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4158 int src_nid, dst_nid;
4160 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4163 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4166 src_nid = cpu_to_node(env->src_cpu);
4167 dst_nid = cpu_to_node(env->dst_cpu);
4169 if (src_nid == dst_nid ||
4170 p->numa_migrate_seq >= sysctl_numa_balancing_settle_count)
4173 if (p->numa_faults[dst_nid] < p->numa_faults[src_nid])
4180 static inline bool migrate_improves_locality(struct task_struct *p,
4186 static inline bool migrate_degrades_locality(struct task_struct *p,
4194 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4197 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4199 int tsk_cache_hot = 0;
4201 * We do not migrate tasks that are:
4202 * 1) throttled_lb_pair, or
4203 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4204 * 3) running (obviously), or
4205 * 4) are cache-hot on their current CPU.
4207 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4210 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4213 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4215 env->flags |= LBF_SOME_PINNED;
4218 * Remember if this task can be migrated to any other cpu in
4219 * our sched_group. We may want to revisit it if we couldn't
4220 * meet load balance goals by pulling other tasks on src_cpu.
4222 * Also avoid computing new_dst_cpu if we have already computed
4223 * one in current iteration.
4225 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4228 /* Prevent to re-select dst_cpu via env's cpus */
4229 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4230 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4231 env->flags |= LBF_DST_PINNED;
4232 env->new_dst_cpu = cpu;
4240 /* Record that we found atleast one task that could run on dst_cpu */
4241 env->flags &= ~LBF_ALL_PINNED;
4243 if (task_running(env->src_rq, p)) {
4244 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4249 * Aggressive migration if:
4250 * 1) destination numa is preferred
4251 * 2) task is cache cold, or
4252 * 3) too many balance attempts have failed.
4254 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4256 tsk_cache_hot = migrate_degrades_locality(p, env);
4258 if (migrate_improves_locality(p, env)) {
4259 #ifdef CONFIG_SCHEDSTATS
4260 if (tsk_cache_hot) {
4261 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4262 schedstat_inc(p, se.statistics.nr_forced_migrations);
4268 if (!tsk_cache_hot ||
4269 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4271 if (tsk_cache_hot) {
4272 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4273 schedstat_inc(p, se.statistics.nr_forced_migrations);
4279 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4284 * move_one_task tries to move exactly one task from busiest to this_rq, as
4285 * part of active balancing operations within "domain".
4286 * Returns 1 if successful and 0 otherwise.
4288 * Called with both runqueues locked.
4290 static int move_one_task(struct lb_env *env)
4292 struct task_struct *p, *n;
4294 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4295 if (!can_migrate_task(p, env))
4300 * Right now, this is only the second place move_task()
4301 * is called, so we can safely collect move_task()
4302 * stats here rather than inside move_task().
4304 schedstat_inc(env->sd, lb_gained[env->idle]);
4310 static unsigned long task_h_load(struct task_struct *p);
4312 static const unsigned int sched_nr_migrate_break = 32;
4315 * move_tasks tries to move up to imbalance weighted load from busiest to
4316 * this_rq, as part of a balancing operation within domain "sd".
4317 * Returns 1 if successful and 0 otherwise.
4319 * Called with both runqueues locked.
4321 static int move_tasks(struct lb_env *env)
4323 struct list_head *tasks = &env->src_rq->cfs_tasks;
4324 struct task_struct *p;
4328 if (env->imbalance <= 0)
4331 while (!list_empty(tasks)) {
4332 p = list_first_entry(tasks, struct task_struct, se.group_node);
4335 /* We've more or less seen every task there is, call it quits */
4336 if (env->loop > env->loop_max)
4339 /* take a breather every nr_migrate tasks */
4340 if (env->loop > env->loop_break) {
4341 env->loop_break += sched_nr_migrate_break;
4342 env->flags |= LBF_NEED_BREAK;
4346 if (!can_migrate_task(p, env))
4349 load = task_h_load(p);
4351 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4354 if ((load / 2) > env->imbalance)
4359 env->imbalance -= load;
4361 #ifdef CONFIG_PREEMPT
4363 * NEWIDLE balancing is a source of latency, so preemptible
4364 * kernels will stop after the first task is pulled to minimize
4365 * the critical section.
4367 if (env->idle == CPU_NEWLY_IDLE)
4372 * We only want to steal up to the prescribed amount of
4375 if (env->imbalance <= 0)
4380 list_move_tail(&p->se.group_node, tasks);
4384 * Right now, this is one of only two places move_task() is called,
4385 * so we can safely collect move_task() stats here rather than
4386 * inside move_task().
4388 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4393 #ifdef CONFIG_FAIR_GROUP_SCHED
4395 * update tg->load_weight by folding this cpu's load_avg
4397 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4399 struct sched_entity *se = tg->se[cpu];
4400 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4402 /* throttled entities do not contribute to load */
4403 if (throttled_hierarchy(cfs_rq))
4406 update_cfs_rq_blocked_load(cfs_rq, 1);
4409 update_entity_load_avg(se, 1);
4411 * We pivot on our runnable average having decayed to zero for
4412 * list removal. This generally implies that all our children
4413 * have also been removed (modulo rounding error or bandwidth
4414 * control); however, such cases are rare and we can fix these
4417 * TODO: fix up out-of-order children on enqueue.
4419 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4420 list_del_leaf_cfs_rq(cfs_rq);
4422 struct rq *rq = rq_of(cfs_rq);
4423 update_rq_runnable_avg(rq, rq->nr_running);
4427 static void update_blocked_averages(int cpu)
4429 struct rq *rq = cpu_rq(cpu);
4430 struct cfs_rq *cfs_rq;
4431 unsigned long flags;
4433 raw_spin_lock_irqsave(&rq->lock, flags);
4434 update_rq_clock(rq);
4436 * Iterates the task_group tree in a bottom up fashion, see
4437 * list_add_leaf_cfs_rq() for details.
4439 for_each_leaf_cfs_rq(rq, cfs_rq) {
4441 * Note: We may want to consider periodically releasing
4442 * rq->lock about these updates so that creating many task
4443 * groups does not result in continually extending hold time.
4445 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
4448 raw_spin_unlock_irqrestore(&rq->lock, flags);
4452 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
4453 * This needs to be done in a top-down fashion because the load of a child
4454 * group is a fraction of its parents load.
4456 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
4458 struct rq *rq = rq_of(cfs_rq);
4459 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
4460 unsigned long now = jiffies;
4463 if (cfs_rq->last_h_load_update == now)
4466 cfs_rq->h_load_next = NULL;
4467 for_each_sched_entity(se) {
4468 cfs_rq = cfs_rq_of(se);
4469 cfs_rq->h_load_next = se;
4470 if (cfs_rq->last_h_load_update == now)
4475 cfs_rq->h_load = cfs_rq->runnable_load_avg;
4476 cfs_rq->last_h_load_update = now;
4479 while ((se = cfs_rq->h_load_next) != NULL) {
4480 load = cfs_rq->h_load;
4481 load = div64_ul(load * se->avg.load_avg_contrib,
4482 cfs_rq->runnable_load_avg + 1);
4483 cfs_rq = group_cfs_rq(se);
4484 cfs_rq->h_load = load;
4485 cfs_rq->last_h_load_update = now;
4489 static unsigned long task_h_load(struct task_struct *p)
4491 struct cfs_rq *cfs_rq = task_cfs_rq(p);
4493 update_cfs_rq_h_load(cfs_rq);
4494 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
4495 cfs_rq->runnable_load_avg + 1);
4498 static inline void update_blocked_averages(int cpu)
4502 static unsigned long task_h_load(struct task_struct *p)
4504 return p->se.avg.load_avg_contrib;
4508 /********** Helpers for find_busiest_group ************************/
4510 * sg_lb_stats - stats of a sched_group required for load_balancing
4512 struct sg_lb_stats {
4513 unsigned long avg_load; /*Avg load across the CPUs of the group */
4514 unsigned long group_load; /* Total load over the CPUs of the group */
4515 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
4516 unsigned long load_per_task;
4517 unsigned long group_power;
4518 unsigned int sum_nr_running; /* Nr tasks running in the group */
4519 unsigned int group_capacity;
4520 unsigned int idle_cpus;
4521 unsigned int group_weight;
4522 int group_imb; /* Is there an imbalance in the group ? */
4523 int group_has_capacity; /* Is there extra capacity in the group? */
4527 * sd_lb_stats - Structure to store the statistics of a sched_domain
4528 * during load balancing.
4530 struct sd_lb_stats {
4531 struct sched_group *busiest; /* Busiest group in this sd */
4532 struct sched_group *local; /* Local group in this sd */
4533 unsigned long total_load; /* Total load of all groups in sd */
4534 unsigned long total_pwr; /* Total power of all groups in sd */
4535 unsigned long avg_load; /* Average load across all groups in sd */
4537 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
4538 struct sg_lb_stats local_stat; /* Statistics of the local group */
4541 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
4544 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
4545 * local_stat because update_sg_lb_stats() does a full clear/assignment.
4546 * We must however clear busiest_stat::avg_load because
4547 * update_sd_pick_busiest() reads this before assignment.
4549 *sds = (struct sd_lb_stats){
4561 * get_sd_load_idx - Obtain the load index for a given sched domain.
4562 * @sd: The sched_domain whose load_idx is to be obtained.
4563 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4565 * Return: The load index.
4567 static inline int get_sd_load_idx(struct sched_domain *sd,
4568 enum cpu_idle_type idle)
4574 load_idx = sd->busy_idx;
4577 case CPU_NEWLY_IDLE:
4578 load_idx = sd->newidle_idx;
4581 load_idx = sd->idle_idx;
4588 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4590 return SCHED_POWER_SCALE;
4593 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4595 return default_scale_freq_power(sd, cpu);
4598 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4600 unsigned long weight = sd->span_weight;
4601 unsigned long smt_gain = sd->smt_gain;
4608 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4610 return default_scale_smt_power(sd, cpu);
4613 static unsigned long scale_rt_power(int cpu)
4615 struct rq *rq = cpu_rq(cpu);
4616 u64 total, available, age_stamp, avg;
4619 * Since we're reading these variables without serialization make sure
4620 * we read them once before doing sanity checks on them.
4622 age_stamp = ACCESS_ONCE(rq->age_stamp);
4623 avg = ACCESS_ONCE(rq->rt_avg);
4625 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
4627 if (unlikely(total < avg)) {
4628 /* Ensures that power won't end up being negative */
4631 available = total - avg;
4634 if (unlikely((s64)total < SCHED_POWER_SCALE))
4635 total = SCHED_POWER_SCALE;
4637 total >>= SCHED_POWER_SHIFT;
4639 return div_u64(available, total);
4642 static void update_cpu_power(struct sched_domain *sd, int cpu)
4644 unsigned long weight = sd->span_weight;
4645 unsigned long power = SCHED_POWER_SCALE;
4646 struct sched_group *sdg = sd->groups;
4648 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4649 if (sched_feat(ARCH_POWER))
4650 power *= arch_scale_smt_power(sd, cpu);
4652 power *= default_scale_smt_power(sd, cpu);
4654 power >>= SCHED_POWER_SHIFT;
4657 sdg->sgp->power_orig = power;
4659 if (sched_feat(ARCH_POWER))
4660 power *= arch_scale_freq_power(sd, cpu);
4662 power *= default_scale_freq_power(sd, cpu);
4664 power >>= SCHED_POWER_SHIFT;
4666 power *= scale_rt_power(cpu);
4667 power >>= SCHED_POWER_SHIFT;
4672 cpu_rq(cpu)->cpu_power = power;
4673 sdg->sgp->power = power;
4676 void update_group_power(struct sched_domain *sd, int cpu)
4678 struct sched_domain *child = sd->child;
4679 struct sched_group *group, *sdg = sd->groups;
4680 unsigned long power, power_orig;
4681 unsigned long interval;
4683 interval = msecs_to_jiffies(sd->balance_interval);
4684 interval = clamp(interval, 1UL, max_load_balance_interval);
4685 sdg->sgp->next_update = jiffies + interval;
4688 update_cpu_power(sd, cpu);
4692 power_orig = power = 0;
4694 if (child->flags & SD_OVERLAP) {
4696 * SD_OVERLAP domains cannot assume that child groups
4697 * span the current group.
4700 for_each_cpu(cpu, sched_group_cpus(sdg)) {
4701 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
4703 power_orig += sg->sgp->power_orig;
4704 power += sg->sgp->power;
4708 * !SD_OVERLAP domains can assume that child groups
4709 * span the current group.
4712 group = child->groups;
4714 power_orig += group->sgp->power_orig;
4715 power += group->sgp->power;
4716 group = group->next;
4717 } while (group != child->groups);
4720 sdg->sgp->power_orig = power_orig;
4721 sdg->sgp->power = power;
4725 * Try and fix up capacity for tiny siblings, this is needed when
4726 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4727 * which on its own isn't powerful enough.
4729 * See update_sd_pick_busiest() and check_asym_packing().
4732 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4735 * Only siblings can have significantly less than SCHED_POWER_SCALE
4737 if (!(sd->flags & SD_SHARE_CPUPOWER))
4741 * If ~90% of the cpu_power is still there, we're good.
4743 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4750 * Group imbalance indicates (and tries to solve) the problem where balancing
4751 * groups is inadequate due to tsk_cpus_allowed() constraints.
4753 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
4754 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
4757 * { 0 1 2 3 } { 4 5 6 7 }
4760 * If we were to balance group-wise we'd place two tasks in the first group and
4761 * two tasks in the second group. Clearly this is undesired as it will overload
4762 * cpu 3 and leave one of the cpus in the second group unused.
4764 * The current solution to this issue is detecting the skew in the first group
4765 * by noticing the lower domain failed to reach balance and had difficulty
4766 * moving tasks due to affinity constraints.
4768 * When this is so detected; this group becomes a candidate for busiest; see
4769 * update_sd_pick_busiest(). And calculcate_imbalance() and
4770 * find_busiest_group() avoid some of the usual balance conditions to allow it
4771 * to create an effective group imbalance.
4773 * This is a somewhat tricky proposition since the next run might not find the
4774 * group imbalance and decide the groups need to be balanced again. A most
4775 * subtle and fragile situation.
4778 static inline int sg_imbalanced(struct sched_group *group)
4780 return group->sgp->imbalance;
4784 * Compute the group capacity.
4786 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
4787 * first dividing out the smt factor and computing the actual number of cores
4788 * and limit power unit capacity with that.
4790 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
4792 unsigned int capacity, smt, cpus;
4793 unsigned int power, power_orig;
4795 power = group->sgp->power;
4796 power_orig = group->sgp->power_orig;
4797 cpus = group->group_weight;
4799 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
4800 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
4801 capacity = cpus / smt; /* cores */
4803 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
4805 capacity = fix_small_capacity(env->sd, group);
4811 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4812 * @env: The load balancing environment.
4813 * @group: sched_group whose statistics are to be updated.
4814 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4815 * @local_group: Does group contain this_cpu.
4816 * @sgs: variable to hold the statistics for this group.
4818 static inline void update_sg_lb_stats(struct lb_env *env,
4819 struct sched_group *group, int load_idx,
4820 int local_group, struct sg_lb_stats *sgs)
4822 unsigned long nr_running;
4826 memset(sgs, 0, sizeof(*sgs));
4828 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4829 struct rq *rq = cpu_rq(i);
4831 nr_running = rq->nr_running;
4833 /* Bias balancing toward cpus of our domain */
4835 load = target_load(i, load_idx);
4837 load = source_load(i, load_idx);
4839 sgs->group_load += load;
4840 sgs->sum_nr_running += nr_running;
4841 sgs->sum_weighted_load += weighted_cpuload(i);
4846 /* Adjust by relative CPU power of the group */
4847 sgs->group_power = group->sgp->power;
4848 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
4850 if (sgs->sum_nr_running)
4851 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4853 sgs->group_weight = group->group_weight;
4855 sgs->group_imb = sg_imbalanced(group);
4856 sgs->group_capacity = sg_capacity(env, group);
4858 if (sgs->group_capacity > sgs->sum_nr_running)
4859 sgs->group_has_capacity = 1;
4863 * update_sd_pick_busiest - return 1 on busiest group
4864 * @env: The load balancing environment.
4865 * @sds: sched_domain statistics
4866 * @sg: sched_group candidate to be checked for being the busiest
4867 * @sgs: sched_group statistics
4869 * Determine if @sg is a busier group than the previously selected
4872 * Return: %true if @sg is a busier group than the previously selected
4873 * busiest group. %false otherwise.
4875 static bool update_sd_pick_busiest(struct lb_env *env,
4876 struct sd_lb_stats *sds,
4877 struct sched_group *sg,
4878 struct sg_lb_stats *sgs)
4880 if (sgs->avg_load <= sds->busiest_stat.avg_load)
4883 if (sgs->sum_nr_running > sgs->group_capacity)
4890 * ASYM_PACKING needs to move all the work to the lowest
4891 * numbered CPUs in the group, therefore mark all groups
4892 * higher than ourself as busy.
4894 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4895 env->dst_cpu < group_first_cpu(sg)) {
4899 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4907 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4908 * @env: The load balancing environment.
4909 * @balance: Should we balance.
4910 * @sds: variable to hold the statistics for this sched_domain.
4912 static inline void update_sd_lb_stats(struct lb_env *env,
4913 struct sd_lb_stats *sds)
4915 struct sched_domain *child = env->sd->child;
4916 struct sched_group *sg = env->sd->groups;
4917 struct sg_lb_stats tmp_sgs;
4918 int load_idx, prefer_sibling = 0;
4920 if (child && child->flags & SD_PREFER_SIBLING)
4923 load_idx = get_sd_load_idx(env->sd, env->idle);
4926 struct sg_lb_stats *sgs = &tmp_sgs;
4929 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4932 sgs = &sds->local_stat;
4934 if (env->idle != CPU_NEWLY_IDLE ||
4935 time_after_eq(jiffies, sg->sgp->next_update))
4936 update_group_power(env->sd, env->dst_cpu);
4939 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
4945 * In case the child domain prefers tasks go to siblings
4946 * first, lower the sg capacity to one so that we'll try
4947 * and move all the excess tasks away. We lower the capacity
4948 * of a group only if the local group has the capacity to fit
4949 * these excess tasks, i.e. nr_running < group_capacity. The
4950 * extra check prevents the case where you always pull from the
4951 * heaviest group when it is already under-utilized (possible
4952 * with a large weight task outweighs the tasks on the system).
4954 if (prefer_sibling && sds->local &&
4955 sds->local_stat.group_has_capacity)
4956 sgs->group_capacity = min(sgs->group_capacity, 1U);
4958 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
4960 sds->busiest_stat = *sgs;
4964 /* Now, start updating sd_lb_stats */
4965 sds->total_load += sgs->group_load;
4966 sds->total_pwr += sgs->group_power;
4969 } while (sg != env->sd->groups);
4973 * check_asym_packing - Check to see if the group is packed into the
4976 * This is primarily intended to used at the sibling level. Some
4977 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4978 * case of POWER7, it can move to lower SMT modes only when higher
4979 * threads are idle. When in lower SMT modes, the threads will
4980 * perform better since they share less core resources. Hence when we
4981 * have idle threads, we want them to be the higher ones.
4983 * This packing function is run on idle threads. It checks to see if
4984 * the busiest CPU in this domain (core in the P7 case) has a higher
4985 * CPU number than the packing function is being run on. Here we are
4986 * assuming lower CPU number will be equivalent to lower a SMT thread
4989 * Return: 1 when packing is required and a task should be moved to
4990 * this CPU. The amount of the imbalance is returned in *imbalance.
4992 * @env: The load balancing environment.
4993 * @sds: Statistics of the sched_domain which is to be packed
4995 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4999 if (!(env->sd->flags & SD_ASYM_PACKING))
5005 busiest_cpu = group_first_cpu(sds->busiest);
5006 if (env->dst_cpu > busiest_cpu)
5009 env->imbalance = DIV_ROUND_CLOSEST(
5010 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5017 * fix_small_imbalance - Calculate the minor imbalance that exists
5018 * amongst the groups of a sched_domain, during
5020 * @env: The load balancing environment.
5021 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5024 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5026 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5027 unsigned int imbn = 2;
5028 unsigned long scaled_busy_load_per_task;
5029 struct sg_lb_stats *local, *busiest;
5031 local = &sds->local_stat;
5032 busiest = &sds->busiest_stat;
5034 if (!local->sum_nr_running)
5035 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5036 else if (busiest->load_per_task > local->load_per_task)
5039 scaled_busy_load_per_task =
5040 (busiest->load_per_task * SCHED_POWER_SCALE) /
5041 busiest->group_power;
5043 if (busiest->avg_load + scaled_busy_load_per_task >=
5044 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5045 env->imbalance = busiest->load_per_task;
5050 * OK, we don't have enough imbalance to justify moving tasks,
5051 * however we may be able to increase total CPU power used by
5055 pwr_now += busiest->group_power *
5056 min(busiest->load_per_task, busiest->avg_load);
5057 pwr_now += local->group_power *
5058 min(local->load_per_task, local->avg_load);
5059 pwr_now /= SCHED_POWER_SCALE;
5061 /* Amount of load we'd subtract */
5062 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5063 busiest->group_power;
5064 if (busiest->avg_load > tmp) {
5065 pwr_move += busiest->group_power *
5066 min(busiest->load_per_task,
5067 busiest->avg_load - tmp);
5070 /* Amount of load we'd add */
5071 if (busiest->avg_load * busiest->group_power <
5072 busiest->load_per_task * SCHED_POWER_SCALE) {
5073 tmp = (busiest->avg_load * busiest->group_power) /
5076 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5079 pwr_move += local->group_power *
5080 min(local->load_per_task, local->avg_load + tmp);
5081 pwr_move /= SCHED_POWER_SCALE;
5083 /* Move if we gain throughput */
5084 if (pwr_move > pwr_now)
5085 env->imbalance = busiest->load_per_task;
5089 * calculate_imbalance - Calculate the amount of imbalance present within the
5090 * groups of a given sched_domain during load balance.
5091 * @env: load balance environment
5092 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5094 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5096 unsigned long max_pull, load_above_capacity = ~0UL;
5097 struct sg_lb_stats *local, *busiest;
5099 local = &sds->local_stat;
5100 busiest = &sds->busiest_stat;
5102 if (busiest->group_imb) {
5104 * In the group_imb case we cannot rely on group-wide averages
5105 * to ensure cpu-load equilibrium, look at wider averages. XXX
5107 busiest->load_per_task =
5108 min(busiest->load_per_task, sds->avg_load);
5112 * In the presence of smp nice balancing, certain scenarios can have
5113 * max load less than avg load(as we skip the groups at or below
5114 * its cpu_power, while calculating max_load..)
5116 if (busiest->avg_load <= sds->avg_load ||
5117 local->avg_load >= sds->avg_load) {
5119 return fix_small_imbalance(env, sds);
5122 if (!busiest->group_imb) {
5124 * Don't want to pull so many tasks that a group would go idle.
5125 * Except of course for the group_imb case, since then we might
5126 * have to drop below capacity to reach cpu-load equilibrium.
5128 load_above_capacity =
5129 (busiest->sum_nr_running - busiest->group_capacity);
5131 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5132 load_above_capacity /= busiest->group_power;
5136 * We're trying to get all the cpus to the average_load, so we don't
5137 * want to push ourselves above the average load, nor do we wish to
5138 * reduce the max loaded cpu below the average load. At the same time,
5139 * we also don't want to reduce the group load below the group capacity
5140 * (so that we can implement power-savings policies etc). Thus we look
5141 * for the minimum possible imbalance.
5143 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5145 /* How much load to actually move to equalise the imbalance */
5146 env->imbalance = min(
5147 max_pull * busiest->group_power,
5148 (sds->avg_load - local->avg_load) * local->group_power
5149 ) / SCHED_POWER_SCALE;
5152 * if *imbalance is less than the average load per runnable task
5153 * there is no guarantee that any tasks will be moved so we'll have
5154 * a think about bumping its value to force at least one task to be
5157 if (env->imbalance < busiest->load_per_task)
5158 return fix_small_imbalance(env, sds);
5161 /******* find_busiest_group() helpers end here *********************/
5164 * find_busiest_group - Returns the busiest group within the sched_domain
5165 * if there is an imbalance. If there isn't an imbalance, and
5166 * the user has opted for power-savings, it returns a group whose
5167 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5168 * such a group exists.
5170 * Also calculates the amount of weighted load which should be moved
5171 * to restore balance.
5173 * @env: The load balancing environment.
5175 * Return: - The busiest group if imbalance exists.
5176 * - If no imbalance and user has opted for power-savings balance,
5177 * return the least loaded group whose CPUs can be
5178 * put to idle by rebalancing its tasks onto our group.
5180 static struct sched_group *find_busiest_group(struct lb_env *env)
5182 struct sg_lb_stats *local, *busiest;
5183 struct sd_lb_stats sds;
5185 init_sd_lb_stats(&sds);
5188 * Compute the various statistics relavent for load balancing at
5191 update_sd_lb_stats(env, &sds);
5192 local = &sds.local_stat;
5193 busiest = &sds.busiest_stat;
5195 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5196 check_asym_packing(env, &sds))
5199 /* There is no busy sibling group to pull tasks from */
5200 if (!sds.busiest || busiest->sum_nr_running == 0)
5203 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5206 * If the busiest group is imbalanced the below checks don't
5207 * work because they assume all things are equal, which typically
5208 * isn't true due to cpus_allowed constraints and the like.
5210 if (busiest->group_imb)
5213 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5214 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5215 !busiest->group_has_capacity)
5219 * If the local group is more busy than the selected busiest group
5220 * don't try and pull any tasks.
5222 if (local->avg_load >= busiest->avg_load)
5226 * Don't pull any tasks if this group is already above the domain
5229 if (local->avg_load >= sds.avg_load)
5232 if (env->idle == CPU_IDLE) {
5234 * This cpu is idle. If the busiest group load doesn't
5235 * have more tasks than the number of available cpu's and
5236 * there is no imbalance between this and busiest group
5237 * wrt to idle cpu's, it is balanced.
5239 if ((local->idle_cpus < busiest->idle_cpus) &&
5240 busiest->sum_nr_running <= busiest->group_weight)
5244 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5245 * imbalance_pct to be conservative.
5247 if (100 * busiest->avg_load <=
5248 env->sd->imbalance_pct * local->avg_load)
5253 /* Looks like there is an imbalance. Compute it */
5254 calculate_imbalance(env, &sds);
5263 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5265 static struct rq *find_busiest_queue(struct lb_env *env,
5266 struct sched_group *group)
5268 struct rq *busiest = NULL, *rq;
5269 unsigned long busiest_load = 0, busiest_power = 1;
5272 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5273 unsigned long power = power_of(i);
5274 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5279 capacity = fix_small_capacity(env->sd, group);
5282 wl = weighted_cpuload(i);
5285 * When comparing with imbalance, use weighted_cpuload()
5286 * which is not scaled with the cpu power.
5288 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5292 * For the load comparisons with the other cpu's, consider
5293 * the weighted_cpuload() scaled with the cpu power, so that
5294 * the load can be moved away from the cpu that is potentially
5295 * running at a lower capacity.
5297 * Thus we're looking for max(wl_i / power_i), crosswise
5298 * multiplication to rid ourselves of the division works out
5299 * to: wl_i * power_j > wl_j * power_i; where j is our
5302 if (wl * busiest_power > busiest_load * power) {
5304 busiest_power = power;
5313 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5314 * so long as it is large enough.
5316 #define MAX_PINNED_INTERVAL 512
5318 /* Working cpumask for load_balance and load_balance_newidle. */
5319 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5321 static int need_active_balance(struct lb_env *env)
5323 struct sched_domain *sd = env->sd;
5325 if (env->idle == CPU_NEWLY_IDLE) {
5328 * ASYM_PACKING needs to force migrate tasks from busy but
5329 * higher numbered CPUs in order to pack all tasks in the
5330 * lowest numbered CPUs.
5332 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5336 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5339 static int active_load_balance_cpu_stop(void *data);
5341 static int should_we_balance(struct lb_env *env)
5343 struct sched_group *sg = env->sd->groups;
5344 struct cpumask *sg_cpus, *sg_mask;
5345 int cpu, balance_cpu = -1;
5348 * In the newly idle case, we will allow all the cpu's
5349 * to do the newly idle load balance.
5351 if (env->idle == CPU_NEWLY_IDLE)
5354 sg_cpus = sched_group_cpus(sg);
5355 sg_mask = sched_group_mask(sg);
5356 /* Try to find first idle cpu */
5357 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5358 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5365 if (balance_cpu == -1)
5366 balance_cpu = group_balance_cpu(sg);
5369 * First idle cpu or the first cpu(busiest) in this sched group
5370 * is eligible for doing load balancing at this and above domains.
5372 return balance_cpu == env->dst_cpu;
5376 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5377 * tasks if there is an imbalance.
5379 static int load_balance(int this_cpu, struct rq *this_rq,
5380 struct sched_domain *sd, enum cpu_idle_type idle,
5381 int *continue_balancing)
5383 int ld_moved, cur_ld_moved, active_balance = 0;
5384 struct sched_domain *sd_parent = sd->parent;
5385 struct sched_group *group;
5387 unsigned long flags;
5388 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5390 struct lb_env env = {
5392 .dst_cpu = this_cpu,
5394 .dst_grpmask = sched_group_cpus(sd->groups),
5396 .loop_break = sched_nr_migrate_break,
5401 * For NEWLY_IDLE load_balancing, we don't need to consider
5402 * other cpus in our group
5404 if (idle == CPU_NEWLY_IDLE)
5405 env.dst_grpmask = NULL;
5407 cpumask_copy(cpus, cpu_active_mask);
5409 schedstat_inc(sd, lb_count[idle]);
5412 if (!should_we_balance(&env)) {
5413 *continue_balancing = 0;
5417 group = find_busiest_group(&env);
5419 schedstat_inc(sd, lb_nobusyg[idle]);
5423 busiest = find_busiest_queue(&env, group);
5425 schedstat_inc(sd, lb_nobusyq[idle]);
5429 BUG_ON(busiest == env.dst_rq);
5431 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
5434 if (busiest->nr_running > 1) {
5436 * Attempt to move tasks. If find_busiest_group has found
5437 * an imbalance but busiest->nr_running <= 1, the group is
5438 * still unbalanced. ld_moved simply stays zero, so it is
5439 * correctly treated as an imbalance.
5441 env.flags |= LBF_ALL_PINNED;
5442 env.src_cpu = busiest->cpu;
5443 env.src_rq = busiest;
5444 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
5447 local_irq_save(flags);
5448 double_rq_lock(env.dst_rq, busiest);
5451 * cur_ld_moved - load moved in current iteration
5452 * ld_moved - cumulative load moved across iterations
5454 cur_ld_moved = move_tasks(&env);
5455 ld_moved += cur_ld_moved;
5456 double_rq_unlock(env.dst_rq, busiest);
5457 local_irq_restore(flags);
5460 * some other cpu did the load balance for us.
5462 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
5463 resched_cpu(env.dst_cpu);
5465 if (env.flags & LBF_NEED_BREAK) {
5466 env.flags &= ~LBF_NEED_BREAK;
5471 * Revisit (affine) tasks on src_cpu that couldn't be moved to
5472 * us and move them to an alternate dst_cpu in our sched_group
5473 * where they can run. The upper limit on how many times we
5474 * iterate on same src_cpu is dependent on number of cpus in our
5477 * This changes load balance semantics a bit on who can move
5478 * load to a given_cpu. In addition to the given_cpu itself
5479 * (or a ilb_cpu acting on its behalf where given_cpu is
5480 * nohz-idle), we now have balance_cpu in a position to move
5481 * load to given_cpu. In rare situations, this may cause
5482 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
5483 * _independently_ and at _same_ time to move some load to
5484 * given_cpu) causing exceess load to be moved to given_cpu.
5485 * This however should not happen so much in practice and
5486 * moreover subsequent load balance cycles should correct the
5487 * excess load moved.
5489 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
5491 /* Prevent to re-select dst_cpu via env's cpus */
5492 cpumask_clear_cpu(env.dst_cpu, env.cpus);
5494 env.dst_rq = cpu_rq(env.new_dst_cpu);
5495 env.dst_cpu = env.new_dst_cpu;
5496 env.flags &= ~LBF_DST_PINNED;
5498 env.loop_break = sched_nr_migrate_break;
5501 * Go back to "more_balance" rather than "redo" since we
5502 * need to continue with same src_cpu.
5508 * We failed to reach balance because of affinity.
5511 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
5513 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
5514 *group_imbalance = 1;
5515 } else if (*group_imbalance)
5516 *group_imbalance = 0;
5519 /* All tasks on this runqueue were pinned by CPU affinity */
5520 if (unlikely(env.flags & LBF_ALL_PINNED)) {
5521 cpumask_clear_cpu(cpu_of(busiest), cpus);
5522 if (!cpumask_empty(cpus)) {
5524 env.loop_break = sched_nr_migrate_break;
5532 schedstat_inc(sd, lb_failed[idle]);
5534 * Increment the failure counter only on periodic balance.
5535 * We do not want newidle balance, which can be very
5536 * frequent, pollute the failure counter causing
5537 * excessive cache_hot migrations and active balances.
5539 if (idle != CPU_NEWLY_IDLE)
5540 sd->nr_balance_failed++;
5542 if (need_active_balance(&env)) {
5543 raw_spin_lock_irqsave(&busiest->lock, flags);
5545 /* don't kick the active_load_balance_cpu_stop,
5546 * if the curr task on busiest cpu can't be
5549 if (!cpumask_test_cpu(this_cpu,
5550 tsk_cpus_allowed(busiest->curr))) {
5551 raw_spin_unlock_irqrestore(&busiest->lock,
5553 env.flags |= LBF_ALL_PINNED;
5554 goto out_one_pinned;
5558 * ->active_balance synchronizes accesses to
5559 * ->active_balance_work. Once set, it's cleared
5560 * only after active load balance is finished.
5562 if (!busiest->active_balance) {
5563 busiest->active_balance = 1;
5564 busiest->push_cpu = this_cpu;
5567 raw_spin_unlock_irqrestore(&busiest->lock, flags);
5569 if (active_balance) {
5570 stop_one_cpu_nowait(cpu_of(busiest),
5571 active_load_balance_cpu_stop, busiest,
5572 &busiest->active_balance_work);
5576 * We've kicked active balancing, reset the failure
5579 sd->nr_balance_failed = sd->cache_nice_tries+1;
5582 sd->nr_balance_failed = 0;
5584 if (likely(!active_balance)) {
5585 /* We were unbalanced, so reset the balancing interval */
5586 sd->balance_interval = sd->min_interval;
5589 * If we've begun active balancing, start to back off. This
5590 * case may not be covered by the all_pinned logic if there
5591 * is only 1 task on the busy runqueue (because we don't call
5594 if (sd->balance_interval < sd->max_interval)
5595 sd->balance_interval *= 2;
5601 schedstat_inc(sd, lb_balanced[idle]);
5603 sd->nr_balance_failed = 0;
5606 /* tune up the balancing interval */
5607 if (((env.flags & LBF_ALL_PINNED) &&
5608 sd->balance_interval < MAX_PINNED_INTERVAL) ||
5609 (sd->balance_interval < sd->max_interval))
5610 sd->balance_interval *= 2;
5618 * idle_balance is called by schedule() if this_cpu is about to become
5619 * idle. Attempts to pull tasks from other CPUs.
5621 void idle_balance(int this_cpu, struct rq *this_rq)
5623 struct sched_domain *sd;
5624 int pulled_task = 0;
5625 unsigned long next_balance = jiffies + HZ;
5628 this_rq->idle_stamp = rq_clock(this_rq);
5630 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5634 * Drop the rq->lock, but keep IRQ/preempt disabled.
5636 raw_spin_unlock(&this_rq->lock);
5638 update_blocked_averages(this_cpu);
5640 for_each_domain(this_cpu, sd) {
5641 unsigned long interval;
5642 int continue_balancing = 1;
5643 u64 t0, domain_cost;
5645 if (!(sd->flags & SD_LOAD_BALANCE))
5648 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
5651 if (sd->flags & SD_BALANCE_NEWIDLE) {
5652 t0 = sched_clock_cpu(this_cpu);
5654 /* If we've pulled tasks over stop searching: */
5655 pulled_task = load_balance(this_cpu, this_rq,
5657 &continue_balancing);
5659 domain_cost = sched_clock_cpu(this_cpu) - t0;
5660 if (domain_cost > sd->max_newidle_lb_cost)
5661 sd->max_newidle_lb_cost = domain_cost;
5663 curr_cost += domain_cost;
5666 interval = msecs_to_jiffies(sd->balance_interval);
5667 if (time_after(next_balance, sd->last_balance + interval))
5668 next_balance = sd->last_balance + interval;
5670 this_rq->idle_stamp = 0;
5676 raw_spin_lock(&this_rq->lock);
5678 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5680 * We are going idle. next_balance may be set based on
5681 * a busy processor. So reset next_balance.
5683 this_rq->next_balance = next_balance;
5686 if (curr_cost > this_rq->max_idle_balance_cost)
5687 this_rq->max_idle_balance_cost = curr_cost;
5691 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5692 * running tasks off the busiest CPU onto idle CPUs. It requires at
5693 * least 1 task to be running on each physical CPU where possible, and
5694 * avoids physical / logical imbalances.
5696 static int active_load_balance_cpu_stop(void *data)
5698 struct rq *busiest_rq = data;
5699 int busiest_cpu = cpu_of(busiest_rq);
5700 int target_cpu = busiest_rq->push_cpu;
5701 struct rq *target_rq = cpu_rq(target_cpu);
5702 struct sched_domain *sd;
5704 raw_spin_lock_irq(&busiest_rq->lock);
5706 /* make sure the requested cpu hasn't gone down in the meantime */
5707 if (unlikely(busiest_cpu != smp_processor_id() ||
5708 !busiest_rq->active_balance))
5711 /* Is there any task to move? */
5712 if (busiest_rq->nr_running <= 1)
5716 * This condition is "impossible", if it occurs
5717 * we need to fix it. Originally reported by
5718 * Bjorn Helgaas on a 128-cpu setup.
5720 BUG_ON(busiest_rq == target_rq);
5722 /* move a task from busiest_rq to target_rq */
5723 double_lock_balance(busiest_rq, target_rq);
5725 /* Search for an sd spanning us and the target CPU. */
5727 for_each_domain(target_cpu, sd) {
5728 if ((sd->flags & SD_LOAD_BALANCE) &&
5729 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5734 struct lb_env env = {
5736 .dst_cpu = target_cpu,
5737 .dst_rq = target_rq,
5738 .src_cpu = busiest_rq->cpu,
5739 .src_rq = busiest_rq,
5743 schedstat_inc(sd, alb_count);
5745 if (move_one_task(&env))
5746 schedstat_inc(sd, alb_pushed);
5748 schedstat_inc(sd, alb_failed);
5751 double_unlock_balance(busiest_rq, target_rq);
5753 busiest_rq->active_balance = 0;
5754 raw_spin_unlock_irq(&busiest_rq->lock);
5758 #ifdef CONFIG_NO_HZ_COMMON
5760 * idle load balancing details
5761 * - When one of the busy CPUs notice that there may be an idle rebalancing
5762 * needed, they will kick the idle load balancer, which then does idle
5763 * load balancing for all the idle CPUs.
5766 cpumask_var_t idle_cpus_mask;
5768 unsigned long next_balance; /* in jiffy units */
5769 } nohz ____cacheline_aligned;
5771 static inline int find_new_ilb(int call_cpu)
5773 int ilb = cpumask_first(nohz.idle_cpus_mask);
5775 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5782 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5783 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5784 * CPU (if there is one).
5786 static void nohz_balancer_kick(int cpu)
5790 nohz.next_balance++;
5792 ilb_cpu = find_new_ilb(cpu);
5794 if (ilb_cpu >= nr_cpu_ids)
5797 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5800 * Use smp_send_reschedule() instead of resched_cpu().
5801 * This way we generate a sched IPI on the target cpu which
5802 * is idle. And the softirq performing nohz idle load balance
5803 * will be run before returning from the IPI.
5805 smp_send_reschedule(ilb_cpu);
5809 static inline void nohz_balance_exit_idle(int cpu)
5811 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5812 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5813 atomic_dec(&nohz.nr_cpus);
5814 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5818 static inline void set_cpu_sd_state_busy(void)
5820 struct sched_domain *sd;
5823 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5825 if (!sd || !sd->nohz_idle)
5829 for (; sd; sd = sd->parent)
5830 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5835 void set_cpu_sd_state_idle(void)
5837 struct sched_domain *sd;
5840 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
5842 if (!sd || sd->nohz_idle)
5846 for (; sd; sd = sd->parent)
5847 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5853 * This routine will record that the cpu is going idle with tick stopped.
5854 * This info will be used in performing idle load balancing in the future.
5856 void nohz_balance_enter_idle(int cpu)
5859 * If this cpu is going down, then nothing needs to be done.
5861 if (!cpu_active(cpu))
5864 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5867 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5868 atomic_inc(&nohz.nr_cpus);
5869 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5872 static int sched_ilb_notifier(struct notifier_block *nfb,
5873 unsigned long action, void *hcpu)
5875 switch (action & ~CPU_TASKS_FROZEN) {
5877 nohz_balance_exit_idle(smp_processor_id());
5885 static DEFINE_SPINLOCK(balancing);
5888 * Scale the max load_balance interval with the number of CPUs in the system.
5889 * This trades load-balance latency on larger machines for less cross talk.
5891 void update_max_interval(void)
5893 max_load_balance_interval = HZ*num_online_cpus()/10;
5897 * It checks each scheduling domain to see if it is due to be balanced,
5898 * and initiates a balancing operation if so.
5900 * Balancing parameters are set up in init_sched_domains.
5902 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5904 int continue_balancing = 1;
5905 struct rq *rq = cpu_rq(cpu);
5906 unsigned long interval;
5907 struct sched_domain *sd;
5908 /* Earliest time when we have to do rebalance again */
5909 unsigned long next_balance = jiffies + 60*HZ;
5910 int update_next_balance = 0;
5911 int need_serialize, need_decay = 0;
5914 update_blocked_averages(cpu);
5917 for_each_domain(cpu, sd) {
5919 * Decay the newidle max times here because this is a regular
5920 * visit to all the domains. Decay ~1% per second.
5922 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
5923 sd->max_newidle_lb_cost =
5924 (sd->max_newidle_lb_cost * 253) / 256;
5925 sd->next_decay_max_lb_cost = jiffies + HZ;
5928 max_cost += sd->max_newidle_lb_cost;
5930 if (!(sd->flags & SD_LOAD_BALANCE))
5934 * Stop the load balance at this level. There is another
5935 * CPU in our sched group which is doing load balancing more
5938 if (!continue_balancing) {
5944 interval = sd->balance_interval;
5945 if (idle != CPU_IDLE)
5946 interval *= sd->busy_factor;
5948 /* scale ms to jiffies */
5949 interval = msecs_to_jiffies(interval);
5950 interval = clamp(interval, 1UL, max_load_balance_interval);
5952 need_serialize = sd->flags & SD_SERIALIZE;
5954 if (need_serialize) {
5955 if (!spin_trylock(&balancing))
5959 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5960 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
5962 * The LBF_DST_PINNED logic could have changed
5963 * env->dst_cpu, so we can't know our idle
5964 * state even if we migrated tasks. Update it.
5966 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
5968 sd->last_balance = jiffies;
5971 spin_unlock(&balancing);
5973 if (time_after(next_balance, sd->last_balance + interval)) {
5974 next_balance = sd->last_balance + interval;
5975 update_next_balance = 1;
5980 * Ensure the rq-wide value also decays but keep it at a
5981 * reasonable floor to avoid funnies with rq->avg_idle.
5983 rq->max_idle_balance_cost =
5984 max((u64)sysctl_sched_migration_cost, max_cost);
5989 * next_balance will be updated only when there is a need.
5990 * When the cpu is attached to null domain for ex, it will not be
5993 if (likely(update_next_balance))
5994 rq->next_balance = next_balance;
5997 #ifdef CONFIG_NO_HZ_COMMON
5999 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6000 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6002 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6004 struct rq *this_rq = cpu_rq(this_cpu);
6008 if (idle != CPU_IDLE ||
6009 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6012 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6013 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6017 * If this cpu gets work to do, stop the load balancing
6018 * work being done for other cpus. Next load
6019 * balancing owner will pick it up.
6024 rq = cpu_rq(balance_cpu);
6026 raw_spin_lock_irq(&rq->lock);
6027 update_rq_clock(rq);
6028 update_idle_cpu_load(rq);
6029 raw_spin_unlock_irq(&rq->lock);
6031 rebalance_domains(balance_cpu, CPU_IDLE);
6033 if (time_after(this_rq->next_balance, rq->next_balance))
6034 this_rq->next_balance = rq->next_balance;
6036 nohz.next_balance = this_rq->next_balance;
6038 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6042 * Current heuristic for kicking the idle load balancer in the presence
6043 * of an idle cpu is the system.
6044 * - This rq has more than one task.
6045 * - At any scheduler domain level, this cpu's scheduler group has multiple
6046 * busy cpu's exceeding the group's power.
6047 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6048 * domain span are idle.
6050 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6052 unsigned long now = jiffies;
6053 struct sched_domain *sd;
6055 if (unlikely(idle_cpu(cpu)))
6059 * We may be recently in ticked or tickless idle mode. At the first
6060 * busy tick after returning from idle, we will update the busy stats.
6062 set_cpu_sd_state_busy();
6063 nohz_balance_exit_idle(cpu);
6066 * None are in tickless mode and hence no need for NOHZ idle load
6069 if (likely(!atomic_read(&nohz.nr_cpus)))
6072 if (time_before(now, nohz.next_balance))
6075 if (rq->nr_running >= 2)
6079 for_each_domain(cpu, sd) {
6080 struct sched_group *sg = sd->groups;
6081 struct sched_group_power *sgp = sg->sgp;
6082 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6084 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6085 goto need_kick_unlock;
6087 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6088 && (cpumask_first_and(nohz.idle_cpus_mask,
6089 sched_domain_span(sd)) < cpu))
6090 goto need_kick_unlock;
6092 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6104 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6108 * run_rebalance_domains is triggered when needed from the scheduler tick.
6109 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6111 static void run_rebalance_domains(struct softirq_action *h)
6113 int this_cpu = smp_processor_id();
6114 struct rq *this_rq = cpu_rq(this_cpu);
6115 enum cpu_idle_type idle = this_rq->idle_balance ?
6116 CPU_IDLE : CPU_NOT_IDLE;
6118 rebalance_domains(this_cpu, idle);
6121 * If this cpu has a pending nohz_balance_kick, then do the
6122 * balancing on behalf of the other idle cpus whose ticks are
6125 nohz_idle_balance(this_cpu, idle);
6128 static inline int on_null_domain(int cpu)
6130 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6134 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6136 void trigger_load_balance(struct rq *rq, int cpu)
6138 /* Don't need to rebalance while attached to NULL domain */
6139 if (time_after_eq(jiffies, rq->next_balance) &&
6140 likely(!on_null_domain(cpu)))
6141 raise_softirq(SCHED_SOFTIRQ);
6142 #ifdef CONFIG_NO_HZ_COMMON
6143 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6144 nohz_balancer_kick(cpu);
6148 static void rq_online_fair(struct rq *rq)
6153 static void rq_offline_fair(struct rq *rq)
6157 /* Ensure any throttled groups are reachable by pick_next_task */
6158 unthrottle_offline_cfs_rqs(rq);
6161 #endif /* CONFIG_SMP */
6164 * scheduler tick hitting a task of our scheduling class:
6166 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6168 struct cfs_rq *cfs_rq;
6169 struct sched_entity *se = &curr->se;
6171 for_each_sched_entity(se) {
6172 cfs_rq = cfs_rq_of(se);
6173 entity_tick(cfs_rq, se, queued);
6176 if (numabalancing_enabled)
6177 task_tick_numa(rq, curr);
6179 update_rq_runnable_avg(rq, 1);
6183 * called on fork with the child task as argument from the parent's context
6184 * - child not yet on the tasklist
6185 * - preemption disabled
6187 static void task_fork_fair(struct task_struct *p)
6189 struct cfs_rq *cfs_rq;
6190 struct sched_entity *se = &p->se, *curr;
6191 int this_cpu = smp_processor_id();
6192 struct rq *rq = this_rq();
6193 unsigned long flags;
6195 raw_spin_lock_irqsave(&rq->lock, flags);
6197 update_rq_clock(rq);
6199 cfs_rq = task_cfs_rq(current);
6200 curr = cfs_rq->curr;
6203 * Not only the cpu but also the task_group of the parent might have
6204 * been changed after parent->se.parent,cfs_rq were copied to
6205 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6206 * of child point to valid ones.
6209 __set_task_cpu(p, this_cpu);
6212 update_curr(cfs_rq);
6215 se->vruntime = curr->vruntime;
6216 place_entity(cfs_rq, se, 1);
6218 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6220 * Upon rescheduling, sched_class::put_prev_task() will place
6221 * 'current' within the tree based on its new key value.
6223 swap(curr->vruntime, se->vruntime);
6224 resched_task(rq->curr);
6227 se->vruntime -= cfs_rq->min_vruntime;
6229 raw_spin_unlock_irqrestore(&rq->lock, flags);
6233 * Priority of the task has changed. Check to see if we preempt
6237 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6243 * Reschedule if we are currently running on this runqueue and
6244 * our priority decreased, or if we are not currently running on
6245 * this runqueue and our priority is higher than the current's
6247 if (rq->curr == p) {
6248 if (p->prio > oldprio)
6249 resched_task(rq->curr);
6251 check_preempt_curr(rq, p, 0);
6254 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6256 struct sched_entity *se = &p->se;
6257 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6260 * Ensure the task's vruntime is normalized, so that when its
6261 * switched back to the fair class the enqueue_entity(.flags=0) will
6262 * do the right thing.
6264 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6265 * have normalized the vruntime, if it was !on_rq, then only when
6266 * the task is sleeping will it still have non-normalized vruntime.
6268 if (!se->on_rq && p->state != TASK_RUNNING) {
6270 * Fix up our vruntime so that the current sleep doesn't
6271 * cause 'unlimited' sleep bonus.
6273 place_entity(cfs_rq, se, 0);
6274 se->vruntime -= cfs_rq->min_vruntime;
6279 * Remove our load from contribution when we leave sched_fair
6280 * and ensure we don't carry in an old decay_count if we
6283 if (se->avg.decay_count) {
6284 __synchronize_entity_decay(se);
6285 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6291 * We switched to the sched_fair class.
6293 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6299 * We were most likely switched from sched_rt, so
6300 * kick off the schedule if running, otherwise just see
6301 * if we can still preempt the current task.
6304 resched_task(rq->curr);
6306 check_preempt_curr(rq, p, 0);
6309 /* Account for a task changing its policy or group.
6311 * This routine is mostly called to set cfs_rq->curr field when a task
6312 * migrates between groups/classes.
6314 static void set_curr_task_fair(struct rq *rq)
6316 struct sched_entity *se = &rq->curr->se;
6318 for_each_sched_entity(se) {
6319 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6321 set_next_entity(cfs_rq, se);
6322 /* ensure bandwidth has been allocated on our new cfs_rq */
6323 account_cfs_rq_runtime(cfs_rq, 0);
6327 void init_cfs_rq(struct cfs_rq *cfs_rq)
6329 cfs_rq->tasks_timeline = RB_ROOT;
6330 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6331 #ifndef CONFIG_64BIT
6332 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6335 atomic64_set(&cfs_rq->decay_counter, 1);
6336 atomic_long_set(&cfs_rq->removed_load, 0);
6340 #ifdef CONFIG_FAIR_GROUP_SCHED
6341 static void task_move_group_fair(struct task_struct *p, int on_rq)
6343 struct cfs_rq *cfs_rq;
6345 * If the task was not on the rq at the time of this cgroup movement
6346 * it must have been asleep, sleeping tasks keep their ->vruntime
6347 * absolute on their old rq until wakeup (needed for the fair sleeper
6348 * bonus in place_entity()).
6350 * If it was on the rq, we've just 'preempted' it, which does convert
6351 * ->vruntime to a relative base.
6353 * Make sure both cases convert their relative position when migrating
6354 * to another cgroup's rq. This does somewhat interfere with the
6355 * fair sleeper stuff for the first placement, but who cares.
6358 * When !on_rq, vruntime of the task has usually NOT been normalized.
6359 * But there are some cases where it has already been normalized:
6361 * - Moving a forked child which is waiting for being woken up by
6362 * wake_up_new_task().
6363 * - Moving a task which has been woken up by try_to_wake_up() and
6364 * waiting for actually being woken up by sched_ttwu_pending().
6366 * To prevent boost or penalty in the new cfs_rq caused by delta
6367 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6369 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6373 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6374 set_task_rq(p, task_cpu(p));
6376 cfs_rq = cfs_rq_of(&p->se);
6377 p->se.vruntime += cfs_rq->min_vruntime;
6380 * migrate_task_rq_fair() will have removed our previous
6381 * contribution, but we must synchronize for ongoing future
6384 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6385 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6390 void free_fair_sched_group(struct task_group *tg)
6394 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6396 for_each_possible_cpu(i) {
6398 kfree(tg->cfs_rq[i]);
6407 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6409 struct cfs_rq *cfs_rq;
6410 struct sched_entity *se;
6413 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6416 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6420 tg->shares = NICE_0_LOAD;
6422 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6424 for_each_possible_cpu(i) {
6425 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6426 GFP_KERNEL, cpu_to_node(i));
6430 se = kzalloc_node(sizeof(struct sched_entity),
6431 GFP_KERNEL, cpu_to_node(i));
6435 init_cfs_rq(cfs_rq);
6436 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
6447 void unregister_fair_sched_group(struct task_group *tg, int cpu)
6449 struct rq *rq = cpu_rq(cpu);
6450 unsigned long flags;
6453 * Only empty task groups can be destroyed; so we can speculatively
6454 * check on_list without danger of it being re-added.
6456 if (!tg->cfs_rq[cpu]->on_list)
6459 raw_spin_lock_irqsave(&rq->lock, flags);
6460 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
6461 raw_spin_unlock_irqrestore(&rq->lock, flags);
6464 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
6465 struct sched_entity *se, int cpu,
6466 struct sched_entity *parent)
6468 struct rq *rq = cpu_rq(cpu);
6472 init_cfs_rq_runtime(cfs_rq);
6474 tg->cfs_rq[cpu] = cfs_rq;
6477 /* se could be NULL for root_task_group */
6482 se->cfs_rq = &rq->cfs;
6484 se->cfs_rq = parent->my_q;
6487 update_load_set(&se->load, 0);
6488 se->parent = parent;
6491 static DEFINE_MUTEX(shares_mutex);
6493 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
6496 unsigned long flags;
6499 * We can't change the weight of the root cgroup.
6504 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
6506 mutex_lock(&shares_mutex);
6507 if (tg->shares == shares)
6510 tg->shares = shares;
6511 for_each_possible_cpu(i) {
6512 struct rq *rq = cpu_rq(i);
6513 struct sched_entity *se;
6516 /* Propagate contribution to hierarchy */
6517 raw_spin_lock_irqsave(&rq->lock, flags);
6519 /* Possible calls to update_curr() need rq clock */
6520 update_rq_clock(rq);
6521 for_each_sched_entity(se)
6522 update_cfs_shares(group_cfs_rq(se));
6523 raw_spin_unlock_irqrestore(&rq->lock, flags);
6527 mutex_unlock(&shares_mutex);
6530 #else /* CONFIG_FAIR_GROUP_SCHED */
6532 void free_fair_sched_group(struct task_group *tg) { }
6534 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6539 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
6541 #endif /* CONFIG_FAIR_GROUP_SCHED */
6544 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
6546 struct sched_entity *se = &task->se;
6547 unsigned int rr_interval = 0;
6550 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
6553 if (rq->cfs.load.weight)
6554 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
6560 * All the scheduling class methods:
6562 const struct sched_class fair_sched_class = {
6563 .next = &idle_sched_class,
6564 .enqueue_task = enqueue_task_fair,
6565 .dequeue_task = dequeue_task_fair,
6566 .yield_task = yield_task_fair,
6567 .yield_to_task = yield_to_task_fair,
6569 .check_preempt_curr = check_preempt_wakeup,
6571 .pick_next_task = pick_next_task_fair,
6572 .put_prev_task = put_prev_task_fair,
6575 .select_task_rq = select_task_rq_fair,
6576 .migrate_task_rq = migrate_task_rq_fair,
6578 .rq_online = rq_online_fair,
6579 .rq_offline = rq_offline_fair,
6581 .task_waking = task_waking_fair,
6584 .set_curr_task = set_curr_task_fair,
6585 .task_tick = task_tick_fair,
6586 .task_fork = task_fork_fair,
6588 .prio_changed = prio_changed_fair,
6589 .switched_from = switched_from_fair,
6590 .switched_to = switched_to_fair,
6592 .get_rr_interval = get_rr_interval_fair,
6594 #ifdef CONFIG_FAIR_GROUP_SCHED
6595 .task_move_group = task_move_group_fair,
6599 #ifdef CONFIG_SCHED_DEBUG
6600 void print_cfs_stats(struct seq_file *m, int cpu)
6602 struct cfs_rq *cfs_rq;
6605 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
6606 print_cfs_rq(m, cpu, cfs_rq);
6611 __init void init_sched_fair_class(void)
6614 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
6616 #ifdef CONFIG_NO_HZ_COMMON
6617 nohz.next_balance = jiffies;
6618 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
6619 cpu_notifier(sched_ilb_notifier, 0);