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 unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
830 /* Portion of address space to scan in MB */
831 unsigned int sysctl_numa_balancing_scan_size = 256;
833 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
834 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 * After skipping a page migration on a shared page, skip N more numa page
838 * migrations unconditionally. This reduces the number of NUMA migrations
839 * in shared memory workloads, and has the effect of pulling tasks towards
840 * where their memory lives, over pulling the memory towards the task.
842 unsigned int sysctl_numa_balancing_migrate_deferred = 16;
844 static unsigned int task_nr_scan_windows(struct task_struct *p)
846 unsigned long rss = 0;
847 unsigned long nr_scan_pages;
850 * Calculations based on RSS as non-present and empty pages are skipped
851 * by the PTE scanner and NUMA hinting faults should be trapped based
854 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
855 rss = get_mm_rss(p->mm);
859 rss = round_up(rss, nr_scan_pages);
860 return rss / nr_scan_pages;
863 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
864 #define MAX_SCAN_WINDOW 2560
866 static unsigned int task_scan_min(struct task_struct *p)
868 unsigned int scan, floor;
869 unsigned int windows = 1;
871 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
872 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
873 floor = 1000 / windows;
875 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
876 return max_t(unsigned int, floor, scan);
879 static unsigned int task_scan_max(struct task_struct *p)
881 unsigned int smin = task_scan_min(p);
884 /* Watch for min being lower than max due to floor calculations */
885 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
886 return max(smin, smax);
890 * Once a preferred node is selected the scheduler balancer will prefer moving
891 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
892 * scans. This will give the process the chance to accumulate more faults on
893 * the preferred node but still allow the scheduler to move the task again if
894 * the nodes CPUs are overloaded.
896 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
898 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
900 rq->nr_numa_running += (p->numa_preferred_nid != -1);
901 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
904 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
906 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
907 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
913 spinlock_t lock; /* nr_tasks, tasks */
916 struct list_head task_list;
919 unsigned long total_faults;
920 unsigned long faults[0];
923 pid_t task_numa_group_id(struct task_struct *p)
925 return p->numa_group ? p->numa_group->gid : 0;
928 static inline int task_faults_idx(int nid, int priv)
930 return 2 * nid + priv;
933 static inline unsigned long task_faults(struct task_struct *p, int nid)
938 return p->numa_faults[task_faults_idx(nid, 0)] +
939 p->numa_faults[task_faults_idx(nid, 1)];
942 static inline unsigned long group_faults(struct task_struct *p, int nid)
947 return p->numa_group->faults[2*nid] + p->numa_group->faults[2*nid+1];
951 * These return the fraction of accesses done by a particular task, or
952 * task group, on a particular numa node. The group weight is given a
953 * larger multiplier, in order to group tasks together that are almost
954 * evenly spread out between numa nodes.
956 static inline unsigned long task_weight(struct task_struct *p, int nid)
958 unsigned long total_faults;
963 total_faults = p->total_numa_faults;
968 return 1000 * task_faults(p, nid) / total_faults;
971 static inline unsigned long group_weight(struct task_struct *p, int nid)
973 if (!p->numa_group || !p->numa_group->total_faults)
976 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
979 static unsigned long weighted_cpuload(const int cpu);
980 static unsigned long source_load(int cpu, int type);
981 static unsigned long target_load(int cpu, int type);
982 static unsigned long power_of(int cpu);
983 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
985 /* Cached statistics for all CPUs within a node */
987 unsigned long nr_running;
990 /* Total compute capacity of CPUs on a node */
993 /* Approximate capacity in terms of runnable tasks on a node */
994 unsigned long capacity;
999 * XXX borrowed from update_sg_lb_stats
1001 static void update_numa_stats(struct numa_stats *ns, int nid)
1005 memset(ns, 0, sizeof(*ns));
1006 for_each_cpu(cpu, cpumask_of_node(nid)) {
1007 struct rq *rq = cpu_rq(cpu);
1009 ns->nr_running += rq->nr_running;
1010 ns->load += weighted_cpuload(cpu);
1011 ns->power += power_of(cpu);
1014 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1015 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1016 ns->has_capacity = (ns->nr_running < ns->capacity);
1019 struct task_numa_env {
1020 struct task_struct *p;
1022 int src_cpu, src_nid;
1023 int dst_cpu, dst_nid;
1025 struct numa_stats src_stats, dst_stats;
1027 int imbalance_pct, idx;
1029 struct task_struct *best_task;
1034 static void task_numa_assign(struct task_numa_env *env,
1035 struct task_struct *p, long imp)
1038 put_task_struct(env->best_task);
1043 env->best_imp = imp;
1044 env->best_cpu = env->dst_cpu;
1048 * This checks if the overall compute and NUMA accesses of the system would
1049 * be improved if the source tasks was migrated to the target dst_cpu taking
1050 * into account that it might be best if task running on the dst_cpu should
1051 * be exchanged with the source task
1053 static void task_numa_compare(struct task_numa_env *env,
1054 long taskimp, long groupimp)
1056 struct rq *src_rq = cpu_rq(env->src_cpu);
1057 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1058 struct task_struct *cur;
1059 long dst_load, src_load;
1061 long imp = (groupimp > 0) ? groupimp : taskimp;
1064 cur = ACCESS_ONCE(dst_rq->curr);
1065 if (cur->pid == 0) /* idle */
1069 * "imp" is the fault differential for the source task between the
1070 * source and destination node. Calculate the total differential for
1071 * the source task and potential destination task. The more negative
1072 * the value is, the more rmeote accesses that would be expected to
1073 * be incurred if the tasks were swapped.
1076 /* Skip this swap candidate if cannot move to the source cpu */
1077 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1081 * If dst and source tasks are in the same NUMA group, or not
1082 * in any group then look only at task weights.
1084 if (cur->numa_group == env->p->numa_group) {
1085 imp = taskimp + task_weight(cur, env->src_nid) -
1086 task_weight(cur, env->dst_nid);
1088 * Add some hysteresis to prevent swapping the
1089 * tasks within a group over tiny differences.
1091 if (cur->numa_group)
1095 * Compare the group weights. If a task is all by
1096 * itself (not part of a group), use the task weight
1099 if (env->p->numa_group)
1104 if (cur->numa_group)
1105 imp += group_weight(cur, env->src_nid) -
1106 group_weight(cur, env->dst_nid);
1108 imp += task_weight(cur, env->src_nid) -
1109 task_weight(cur, env->dst_nid);
1113 if (imp < env->best_imp)
1117 /* Is there capacity at our destination? */
1118 if (env->src_stats.has_capacity &&
1119 !env->dst_stats.has_capacity)
1125 /* Balance doesn't matter much if we're running a task per cpu */
1126 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1130 * In the overloaded case, try and keep the load balanced.
1133 dst_load = env->dst_stats.load;
1134 src_load = env->src_stats.load;
1136 /* XXX missing power terms */
1137 load = task_h_load(env->p);
1142 load = task_h_load(cur);
1147 /* make src_load the smaller */
1148 if (dst_load < src_load)
1149 swap(dst_load, src_load);
1151 if (src_load * env->imbalance_pct < dst_load * 100)
1155 task_numa_assign(env, cur, imp);
1160 static void task_numa_find_cpu(struct task_numa_env *env,
1161 long taskimp, long groupimp)
1165 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1166 /* Skip this CPU if the source task cannot migrate */
1167 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1171 task_numa_compare(env, taskimp, groupimp);
1175 static int task_numa_migrate(struct task_struct *p)
1177 struct task_numa_env env = {
1180 .src_cpu = task_cpu(p),
1181 .src_nid = task_node(p),
1183 .imbalance_pct = 112,
1189 struct sched_domain *sd;
1190 unsigned long taskweight, groupweight;
1192 long taskimp, groupimp;
1195 * Pick the lowest SD_NUMA domain, as that would have the smallest
1196 * imbalance and would be the first to start moving tasks about.
1198 * And we want to avoid any moving of tasks about, as that would create
1199 * random movement of tasks -- counter the numa conditions we're trying
1203 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1204 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1207 taskweight = task_weight(p, env.src_nid);
1208 groupweight = group_weight(p, env.src_nid);
1209 update_numa_stats(&env.src_stats, env.src_nid);
1210 env.dst_nid = p->numa_preferred_nid;
1211 taskimp = task_weight(p, env.dst_nid) - taskweight;
1212 groupimp = group_weight(p, env.dst_nid) - groupweight;
1213 update_numa_stats(&env.dst_stats, env.dst_nid);
1215 /* If the preferred nid has capacity, try to use it. */
1216 if (env.dst_stats.has_capacity)
1217 task_numa_find_cpu(&env, taskimp, groupimp);
1219 /* No space available on the preferred nid. Look elsewhere. */
1220 if (env.best_cpu == -1) {
1221 for_each_online_node(nid) {
1222 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1225 /* Only consider nodes where both task and groups benefit */
1226 taskimp = task_weight(p, nid) - taskweight;
1227 groupimp = group_weight(p, nid) - groupweight;
1228 if (taskimp < 0 && groupimp < 0)
1232 update_numa_stats(&env.dst_stats, env.dst_nid);
1233 task_numa_find_cpu(&env, taskimp, groupimp);
1237 /* No better CPU than the current one was found. */
1238 if (env.best_cpu == -1)
1241 sched_setnuma(p, env.dst_nid);
1244 * Reset the scan period if the task is being rescheduled on an
1245 * alternative node to recheck if the tasks is now properly placed.
1247 p->numa_scan_period = task_scan_min(p);
1249 if (env.best_task == NULL) {
1250 int ret = migrate_task_to(p, env.best_cpu);
1254 ret = migrate_swap(p, env.best_task);
1255 put_task_struct(env.best_task);
1259 /* Attempt to migrate a task to a CPU on the preferred node. */
1260 static void numa_migrate_preferred(struct task_struct *p)
1262 /* This task has no NUMA fault statistics yet */
1263 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1266 /* Periodically retry migrating the task to the preferred node */
1267 p->numa_migrate_retry = jiffies + HZ;
1269 /* Success if task is already running on preferred CPU */
1270 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid)
1273 /* Otherwise, try migrate to a CPU on the preferred node */
1274 task_numa_migrate(p);
1278 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1279 * increments. The more local the fault statistics are, the higher the scan
1280 * period will be for the next scan window. If local/remote ratio is below
1281 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1282 * scan period will decrease
1284 #define NUMA_PERIOD_SLOTS 10
1285 #define NUMA_PERIOD_THRESHOLD 3
1288 * Increase the scan period (slow down scanning) if the majority of
1289 * our memory is already on our local node, or if the majority of
1290 * the page accesses are shared with other processes.
1291 * Otherwise, decrease the scan period.
1293 static void update_task_scan_period(struct task_struct *p,
1294 unsigned long shared, unsigned long private)
1296 unsigned int period_slot;
1300 unsigned long remote = p->numa_faults_locality[0];
1301 unsigned long local = p->numa_faults_locality[1];
1304 * If there were no record hinting faults then either the task is
1305 * completely idle or all activity is areas that are not of interest
1306 * to automatic numa balancing. Scan slower
1308 if (local + shared == 0) {
1309 p->numa_scan_period = min(p->numa_scan_period_max,
1310 p->numa_scan_period << 1);
1312 p->mm->numa_next_scan = jiffies +
1313 msecs_to_jiffies(p->numa_scan_period);
1319 * Prepare to scale scan period relative to the current period.
1320 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1321 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1322 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1324 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1325 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1326 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1327 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1330 diff = slot * period_slot;
1332 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1335 * Scale scan rate increases based on sharing. There is an
1336 * inverse relationship between the degree of sharing and
1337 * the adjustment made to the scanning period. Broadly
1338 * speaking the intent is that there is little point
1339 * scanning faster if shared accesses dominate as it may
1340 * simply bounce migrations uselessly
1342 period_slot = DIV_ROUND_UP(diff, NUMA_PERIOD_SLOTS);
1343 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1344 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1347 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1348 task_scan_min(p), task_scan_max(p));
1349 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1352 static void task_numa_placement(struct task_struct *p)
1354 int seq, nid, max_nid = -1, max_group_nid = -1;
1355 unsigned long max_faults = 0, max_group_faults = 0;
1356 unsigned long fault_types[2] = { 0, 0 };
1357 spinlock_t *group_lock = NULL;
1359 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1360 if (p->numa_scan_seq == seq)
1362 p->numa_scan_seq = seq;
1363 p->numa_scan_period_max = task_scan_max(p);
1365 /* If the task is part of a group prevent parallel updates to group stats */
1366 if (p->numa_group) {
1367 group_lock = &p->numa_group->lock;
1368 spin_lock(group_lock);
1371 /* Find the node with the highest number of faults */
1372 for_each_online_node(nid) {
1373 unsigned long faults = 0, group_faults = 0;
1376 for (priv = 0; priv < 2; priv++) {
1379 i = task_faults_idx(nid, priv);
1380 diff = -p->numa_faults[i];
1382 /* Decay existing window, copy faults since last scan */
1383 p->numa_faults[i] >>= 1;
1384 p->numa_faults[i] += p->numa_faults_buffer[i];
1385 fault_types[priv] += p->numa_faults_buffer[i];
1386 p->numa_faults_buffer[i] = 0;
1388 faults += p->numa_faults[i];
1389 diff += p->numa_faults[i];
1390 p->total_numa_faults += diff;
1391 if (p->numa_group) {
1392 /* safe because we can only change our own group */
1393 p->numa_group->faults[i] += diff;
1394 p->numa_group->total_faults += diff;
1395 group_faults += p->numa_group->faults[i];
1399 if (faults > max_faults) {
1400 max_faults = faults;
1404 if (group_faults > max_group_faults) {
1405 max_group_faults = group_faults;
1406 max_group_nid = nid;
1410 update_task_scan_period(p, fault_types[0], fault_types[1]);
1412 if (p->numa_group) {
1414 * If the preferred task and group nids are different,
1415 * iterate over the nodes again to find the best place.
1417 if (max_nid != max_group_nid) {
1418 unsigned long weight, max_weight = 0;
1420 for_each_online_node(nid) {
1421 weight = task_weight(p, nid) + group_weight(p, nid);
1422 if (weight > max_weight) {
1423 max_weight = weight;
1429 spin_unlock(group_lock);
1432 /* Preferred node as the node with the most faults */
1433 if (max_faults && max_nid != p->numa_preferred_nid) {
1434 /* Update the preferred nid and migrate task if possible */
1435 sched_setnuma(p, max_nid);
1436 numa_migrate_preferred(p);
1440 static inline int get_numa_group(struct numa_group *grp)
1442 return atomic_inc_not_zero(&grp->refcount);
1445 static inline void put_numa_group(struct numa_group *grp)
1447 if (atomic_dec_and_test(&grp->refcount))
1448 kfree_rcu(grp, rcu);
1451 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1454 struct numa_group *grp, *my_grp;
1455 struct task_struct *tsk;
1457 int cpu = cpupid_to_cpu(cpupid);
1460 if (unlikely(!p->numa_group)) {
1461 unsigned int size = sizeof(struct numa_group) +
1462 2*nr_node_ids*sizeof(unsigned long);
1464 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1468 atomic_set(&grp->refcount, 1);
1469 spin_lock_init(&grp->lock);
1470 INIT_LIST_HEAD(&grp->task_list);
1473 for (i = 0; i < 2*nr_node_ids; i++)
1474 grp->faults[i] = p->numa_faults[i];
1476 grp->total_faults = p->total_numa_faults;
1478 list_add(&p->numa_entry, &grp->task_list);
1480 rcu_assign_pointer(p->numa_group, grp);
1484 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1486 if (!cpupid_match_pid(tsk, cpupid))
1489 grp = rcu_dereference(tsk->numa_group);
1493 my_grp = p->numa_group;
1498 * Only join the other group if its bigger; if we're the bigger group,
1499 * the other task will join us.
1501 if (my_grp->nr_tasks > grp->nr_tasks)
1505 * Tie-break on the grp address.
1507 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1510 /* Always join threads in the same process. */
1511 if (tsk->mm == current->mm)
1514 /* Simple filter to avoid false positives due to PID collisions */
1515 if (flags & TNF_SHARED)
1518 /* Update priv based on whether false sharing was detected */
1521 if (join && !get_numa_group(grp))
1529 double_lock(&my_grp->lock, &grp->lock);
1531 for (i = 0; i < 2*nr_node_ids; i++) {
1532 my_grp->faults[i] -= p->numa_faults[i];
1533 grp->faults[i] += p->numa_faults[i];
1535 my_grp->total_faults -= p->total_numa_faults;
1536 grp->total_faults += p->total_numa_faults;
1538 list_move(&p->numa_entry, &grp->task_list);
1542 spin_unlock(&my_grp->lock);
1543 spin_unlock(&grp->lock);
1545 rcu_assign_pointer(p->numa_group, grp);
1547 put_numa_group(my_grp);
1555 void task_numa_free(struct task_struct *p)
1557 struct numa_group *grp = p->numa_group;
1559 void *numa_faults = p->numa_faults;
1562 spin_lock(&grp->lock);
1563 for (i = 0; i < 2*nr_node_ids; i++)
1564 grp->faults[i] -= p->numa_faults[i];
1565 grp->total_faults -= p->total_numa_faults;
1567 list_del(&p->numa_entry);
1569 spin_unlock(&grp->lock);
1570 rcu_assign_pointer(p->numa_group, NULL);
1571 put_numa_group(grp);
1574 p->numa_faults = NULL;
1575 p->numa_faults_buffer = NULL;
1580 * Got a PROT_NONE fault for a page on @node.
1582 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1584 struct task_struct *p = current;
1585 bool migrated = flags & TNF_MIGRATED;
1588 if (!numabalancing_enabled)
1591 /* for example, ksmd faulting in a user's mm */
1595 /* Do not worry about placement if exiting */
1596 if (p->state == TASK_DEAD)
1599 /* Allocate buffer to track faults on a per-node basis */
1600 if (unlikely(!p->numa_faults)) {
1601 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1603 /* numa_faults and numa_faults_buffer share the allocation */
1604 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1605 if (!p->numa_faults)
1608 BUG_ON(p->numa_faults_buffer);
1609 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1610 p->total_numa_faults = 0;
1611 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1615 * First accesses are treated as private, otherwise consider accesses
1616 * to be private if the accessing pid has not changed
1618 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1621 priv = cpupid_match_pid(p, last_cpupid);
1622 if (!priv && !(flags & TNF_NO_GROUP))
1623 task_numa_group(p, last_cpupid, flags, &priv);
1626 task_numa_placement(p);
1629 * Retry task to preferred node migration periodically, in case it
1630 * case it previously failed, or the scheduler moved us.
1632 if (time_after(jiffies, p->numa_migrate_retry))
1633 numa_migrate_preferred(p);
1636 p->numa_pages_migrated += pages;
1638 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1639 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1642 static void reset_ptenuma_scan(struct task_struct *p)
1644 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1645 p->mm->numa_scan_offset = 0;
1649 * The expensive part of numa migration is done from task_work context.
1650 * Triggered from task_tick_numa().
1652 void task_numa_work(struct callback_head *work)
1654 unsigned long migrate, next_scan, now = jiffies;
1655 struct task_struct *p = current;
1656 struct mm_struct *mm = p->mm;
1657 struct vm_area_struct *vma;
1658 unsigned long start, end;
1659 unsigned long nr_pte_updates = 0;
1662 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1664 work->next = work; /* protect against double add */
1666 * Who cares about NUMA placement when they're dying.
1668 * NOTE: make sure not to dereference p->mm before this check,
1669 * exit_task_work() happens _after_ exit_mm() so we could be called
1670 * without p->mm even though we still had it when we enqueued this
1673 if (p->flags & PF_EXITING)
1676 if (!mm->numa_next_scan) {
1677 mm->numa_next_scan = now +
1678 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1682 * Enforce maximal scan/migration frequency..
1684 migrate = mm->numa_next_scan;
1685 if (time_before(now, migrate))
1688 if (p->numa_scan_period == 0) {
1689 p->numa_scan_period_max = task_scan_max(p);
1690 p->numa_scan_period = task_scan_min(p);
1693 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1694 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1698 * Delay this task enough that another task of this mm will likely win
1699 * the next time around.
1701 p->node_stamp += 2 * TICK_NSEC;
1703 start = mm->numa_scan_offset;
1704 pages = sysctl_numa_balancing_scan_size;
1705 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1709 down_read(&mm->mmap_sem);
1710 vma = find_vma(mm, start);
1712 reset_ptenuma_scan(p);
1716 for (; vma; vma = vma->vm_next) {
1717 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1721 * Shared library pages mapped by multiple processes are not
1722 * migrated as it is expected they are cache replicated. Avoid
1723 * hinting faults in read-only file-backed mappings or the vdso
1724 * as migrating the pages will be of marginal benefit.
1727 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1731 start = max(start, vma->vm_start);
1732 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1733 end = min(end, vma->vm_end);
1734 nr_pte_updates += change_prot_numa(vma, start, end);
1737 * Scan sysctl_numa_balancing_scan_size but ensure that
1738 * at least one PTE is updated so that unused virtual
1739 * address space is quickly skipped.
1742 pages -= (end - start) >> PAGE_SHIFT;
1747 } while (end != vma->vm_end);
1752 * It is possible to reach the end of the VMA list but the last few
1753 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1754 * would find the !migratable VMA on the next scan but not reset the
1755 * scanner to the start so check it now.
1758 mm->numa_scan_offset = start;
1760 reset_ptenuma_scan(p);
1761 up_read(&mm->mmap_sem);
1765 * Drive the periodic memory faults..
1767 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1769 struct callback_head *work = &curr->numa_work;
1773 * We don't care about NUMA placement if we don't have memory.
1775 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1779 * Using runtime rather than walltime has the dual advantage that
1780 * we (mostly) drive the selection from busy threads and that the
1781 * task needs to have done some actual work before we bother with
1784 now = curr->se.sum_exec_runtime;
1785 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1787 if (now - curr->node_stamp > period) {
1788 if (!curr->node_stamp)
1789 curr->numa_scan_period = task_scan_min(curr);
1790 curr->node_stamp += period;
1792 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1793 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1794 task_work_add(curr, work, true);
1799 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1803 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1807 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1810 #endif /* CONFIG_NUMA_BALANCING */
1813 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1815 update_load_add(&cfs_rq->load, se->load.weight);
1816 if (!parent_entity(se))
1817 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1819 if (entity_is_task(se)) {
1820 struct rq *rq = rq_of(cfs_rq);
1822 account_numa_enqueue(rq, task_of(se));
1823 list_add(&se->group_node, &rq->cfs_tasks);
1826 cfs_rq->nr_running++;
1830 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1832 update_load_sub(&cfs_rq->load, se->load.weight);
1833 if (!parent_entity(se))
1834 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1835 if (entity_is_task(se)) {
1836 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1837 list_del_init(&se->group_node);
1839 cfs_rq->nr_running--;
1842 #ifdef CONFIG_FAIR_GROUP_SCHED
1844 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1849 * Use this CPU's actual weight instead of the last load_contribution
1850 * to gain a more accurate current total weight. See
1851 * update_cfs_rq_load_contribution().
1853 tg_weight = atomic_long_read(&tg->load_avg);
1854 tg_weight -= cfs_rq->tg_load_contrib;
1855 tg_weight += cfs_rq->load.weight;
1860 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1862 long tg_weight, load, shares;
1864 tg_weight = calc_tg_weight(tg, cfs_rq);
1865 load = cfs_rq->load.weight;
1867 shares = (tg->shares * load);
1869 shares /= tg_weight;
1871 if (shares < MIN_SHARES)
1872 shares = MIN_SHARES;
1873 if (shares > tg->shares)
1874 shares = tg->shares;
1878 # else /* CONFIG_SMP */
1879 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1883 # endif /* CONFIG_SMP */
1884 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1885 unsigned long weight)
1888 /* commit outstanding execution time */
1889 if (cfs_rq->curr == se)
1890 update_curr(cfs_rq);
1891 account_entity_dequeue(cfs_rq, se);
1894 update_load_set(&se->load, weight);
1897 account_entity_enqueue(cfs_rq, se);
1900 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1902 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1904 struct task_group *tg;
1905 struct sched_entity *se;
1909 se = tg->se[cpu_of(rq_of(cfs_rq))];
1910 if (!se || throttled_hierarchy(cfs_rq))
1913 if (likely(se->load.weight == tg->shares))
1916 shares = calc_cfs_shares(cfs_rq, tg);
1918 reweight_entity(cfs_rq_of(se), se, shares);
1920 #else /* CONFIG_FAIR_GROUP_SCHED */
1921 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1924 #endif /* CONFIG_FAIR_GROUP_SCHED */
1928 * We choose a half-life close to 1 scheduling period.
1929 * Note: The tables below are dependent on this value.
1931 #define LOAD_AVG_PERIOD 32
1932 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1933 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1935 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1936 static const u32 runnable_avg_yN_inv[] = {
1937 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1938 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1939 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1940 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1941 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1942 0x85aac367, 0x82cd8698,
1946 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1947 * over-estimates when re-combining.
1949 static const u32 runnable_avg_yN_sum[] = {
1950 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1951 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1952 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1957 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1959 static __always_inline u64 decay_load(u64 val, u64 n)
1961 unsigned int local_n;
1965 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1968 /* after bounds checking we can collapse to 32-bit */
1972 * As y^PERIOD = 1/2, we can combine
1973 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1974 * With a look-up table which covers k^n (n<PERIOD)
1976 * To achieve constant time decay_load.
1978 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1979 val >>= local_n / LOAD_AVG_PERIOD;
1980 local_n %= LOAD_AVG_PERIOD;
1983 val *= runnable_avg_yN_inv[local_n];
1984 /* We don't use SRR here since we always want to round down. */
1989 * For updates fully spanning n periods, the contribution to runnable
1990 * average will be: \Sum 1024*y^n
1992 * We can compute this reasonably efficiently by combining:
1993 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1995 static u32 __compute_runnable_contrib(u64 n)
1999 if (likely(n <= LOAD_AVG_PERIOD))
2000 return runnable_avg_yN_sum[n];
2001 else if (unlikely(n >= LOAD_AVG_MAX_N))
2002 return LOAD_AVG_MAX;
2004 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2006 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2007 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2009 n -= LOAD_AVG_PERIOD;
2010 } while (n > LOAD_AVG_PERIOD);
2012 contrib = decay_load(contrib, n);
2013 return contrib + runnable_avg_yN_sum[n];
2017 * We can represent the historical contribution to runnable average as the
2018 * coefficients of a geometric series. To do this we sub-divide our runnable
2019 * history into segments of approximately 1ms (1024us); label the segment that
2020 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2022 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2024 * (now) (~1ms ago) (~2ms ago)
2026 * Let u_i denote the fraction of p_i that the entity was runnable.
2028 * We then designate the fractions u_i as our co-efficients, yielding the
2029 * following representation of historical load:
2030 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2032 * We choose y based on the with of a reasonably scheduling period, fixing:
2035 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2036 * approximately half as much as the contribution to load within the last ms
2039 * When a period "rolls over" and we have new u_0`, multiplying the previous
2040 * sum again by y is sufficient to update:
2041 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2042 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2044 static __always_inline int __update_entity_runnable_avg(u64 now,
2045 struct sched_avg *sa,
2049 u32 runnable_contrib;
2050 int delta_w, decayed = 0;
2052 delta = now - sa->last_runnable_update;
2054 * This should only happen when time goes backwards, which it
2055 * unfortunately does during sched clock init when we swap over to TSC.
2057 if ((s64)delta < 0) {
2058 sa->last_runnable_update = now;
2063 * Use 1024ns as the unit of measurement since it's a reasonable
2064 * approximation of 1us and fast to compute.
2069 sa->last_runnable_update = now;
2071 /* delta_w is the amount already accumulated against our next period */
2072 delta_w = sa->runnable_avg_period % 1024;
2073 if (delta + delta_w >= 1024) {
2074 /* period roll-over */
2078 * Now that we know we're crossing a period boundary, figure
2079 * out how much from delta we need to complete the current
2080 * period and accrue it.
2082 delta_w = 1024 - delta_w;
2084 sa->runnable_avg_sum += delta_w;
2085 sa->runnable_avg_period += delta_w;
2089 /* Figure out how many additional periods this update spans */
2090 periods = delta / 1024;
2093 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2095 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2098 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2099 runnable_contrib = __compute_runnable_contrib(periods);
2101 sa->runnable_avg_sum += runnable_contrib;
2102 sa->runnable_avg_period += runnable_contrib;
2105 /* Remainder of delta accrued against u_0` */
2107 sa->runnable_avg_sum += delta;
2108 sa->runnable_avg_period += delta;
2113 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2114 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2116 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2117 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2119 decays -= se->avg.decay_count;
2123 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2124 se->avg.decay_count = 0;
2129 #ifdef CONFIG_FAIR_GROUP_SCHED
2130 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2133 struct task_group *tg = cfs_rq->tg;
2136 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2137 tg_contrib -= cfs_rq->tg_load_contrib;
2139 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2140 atomic_long_add(tg_contrib, &tg->load_avg);
2141 cfs_rq->tg_load_contrib += tg_contrib;
2146 * Aggregate cfs_rq runnable averages into an equivalent task_group
2147 * representation for computing load contributions.
2149 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2150 struct cfs_rq *cfs_rq)
2152 struct task_group *tg = cfs_rq->tg;
2155 /* The fraction of a cpu used by this cfs_rq */
2156 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2157 sa->runnable_avg_period + 1);
2158 contrib -= cfs_rq->tg_runnable_contrib;
2160 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2161 atomic_add(contrib, &tg->runnable_avg);
2162 cfs_rq->tg_runnable_contrib += contrib;
2166 static inline void __update_group_entity_contrib(struct sched_entity *se)
2168 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2169 struct task_group *tg = cfs_rq->tg;
2174 contrib = cfs_rq->tg_load_contrib * tg->shares;
2175 se->avg.load_avg_contrib = div_u64(contrib,
2176 atomic_long_read(&tg->load_avg) + 1);
2179 * For group entities we need to compute a correction term in the case
2180 * that they are consuming <1 cpu so that we would contribute the same
2181 * load as a task of equal weight.
2183 * Explicitly co-ordinating this measurement would be expensive, but
2184 * fortunately the sum of each cpus contribution forms a usable
2185 * lower-bound on the true value.
2187 * Consider the aggregate of 2 contributions. Either they are disjoint
2188 * (and the sum represents true value) or they are disjoint and we are
2189 * understating by the aggregate of their overlap.
2191 * Extending this to N cpus, for a given overlap, the maximum amount we
2192 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2193 * cpus that overlap for this interval and w_i is the interval width.
2195 * On a small machine; the first term is well-bounded which bounds the
2196 * total error since w_i is a subset of the period. Whereas on a
2197 * larger machine, while this first term can be larger, if w_i is the
2198 * of consequential size guaranteed to see n_i*w_i quickly converge to
2199 * our upper bound of 1-cpu.
2201 runnable_avg = atomic_read(&tg->runnable_avg);
2202 if (runnable_avg < NICE_0_LOAD) {
2203 se->avg.load_avg_contrib *= runnable_avg;
2204 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2208 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2209 int force_update) {}
2210 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2211 struct cfs_rq *cfs_rq) {}
2212 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2215 static inline void __update_task_entity_contrib(struct sched_entity *se)
2219 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2220 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2221 contrib /= (se->avg.runnable_avg_period + 1);
2222 se->avg.load_avg_contrib = scale_load(contrib);
2225 /* Compute the current contribution to load_avg by se, return any delta */
2226 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2228 long old_contrib = se->avg.load_avg_contrib;
2230 if (entity_is_task(se)) {
2231 __update_task_entity_contrib(se);
2233 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2234 __update_group_entity_contrib(se);
2237 return se->avg.load_avg_contrib - old_contrib;
2240 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2243 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2244 cfs_rq->blocked_load_avg -= load_contrib;
2246 cfs_rq->blocked_load_avg = 0;
2249 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2251 /* Update a sched_entity's runnable average */
2252 static inline void update_entity_load_avg(struct sched_entity *se,
2255 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2260 * For a group entity we need to use their owned cfs_rq_clock_task() in
2261 * case they are the parent of a throttled hierarchy.
2263 if (entity_is_task(se))
2264 now = cfs_rq_clock_task(cfs_rq);
2266 now = cfs_rq_clock_task(group_cfs_rq(se));
2268 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2271 contrib_delta = __update_entity_load_avg_contrib(se);
2277 cfs_rq->runnable_load_avg += contrib_delta;
2279 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2283 * Decay the load contributed by all blocked children and account this so that
2284 * their contribution may appropriately discounted when they wake up.
2286 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2288 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2291 decays = now - cfs_rq->last_decay;
2292 if (!decays && !force_update)
2295 if (atomic_long_read(&cfs_rq->removed_load)) {
2296 unsigned long removed_load;
2297 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2298 subtract_blocked_load_contrib(cfs_rq, removed_load);
2302 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2304 atomic64_add(decays, &cfs_rq->decay_counter);
2305 cfs_rq->last_decay = now;
2308 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2311 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2313 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2314 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2317 /* Add the load generated by se into cfs_rq's child load-average */
2318 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2319 struct sched_entity *se,
2323 * We track migrations using entity decay_count <= 0, on a wake-up
2324 * migration we use a negative decay count to track the remote decays
2325 * accumulated while sleeping.
2327 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2328 * are seen by enqueue_entity_load_avg() as a migration with an already
2329 * constructed load_avg_contrib.
2331 if (unlikely(se->avg.decay_count <= 0)) {
2332 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2333 if (se->avg.decay_count) {
2335 * In a wake-up migration we have to approximate the
2336 * time sleeping. This is because we can't synchronize
2337 * clock_task between the two cpus, and it is not
2338 * guaranteed to be read-safe. Instead, we can
2339 * approximate this using our carried decays, which are
2340 * explicitly atomically readable.
2342 se->avg.last_runnable_update -= (-se->avg.decay_count)
2344 update_entity_load_avg(se, 0);
2345 /* Indicate that we're now synchronized and on-rq */
2346 se->avg.decay_count = 0;
2351 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2352 * would have made count negative); we must be careful to avoid
2353 * double-accounting blocked time after synchronizing decays.
2355 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2359 /* migrated tasks did not contribute to our blocked load */
2361 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2362 update_entity_load_avg(se, 0);
2365 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2366 /* we force update consideration on load-balancer moves */
2367 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2371 * Remove se's load from this cfs_rq child load-average, if the entity is
2372 * transitioning to a blocked state we track its projected decay using
2375 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2376 struct sched_entity *se,
2379 update_entity_load_avg(se, 1);
2380 /* we force update consideration on load-balancer moves */
2381 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2383 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2385 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2386 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2387 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2391 * Update the rq's load with the elapsed running time before entering
2392 * idle. if the last scheduled task is not a CFS task, idle_enter will
2393 * be the only way to update the runnable statistic.
2395 void idle_enter_fair(struct rq *this_rq)
2397 update_rq_runnable_avg(this_rq, 1);
2401 * Update the rq's load with the elapsed idle time before a task is
2402 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2403 * be the only way to update the runnable statistic.
2405 void idle_exit_fair(struct rq *this_rq)
2407 update_rq_runnable_avg(this_rq, 0);
2411 static inline void update_entity_load_avg(struct sched_entity *se,
2412 int update_cfs_rq) {}
2413 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2414 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2415 struct sched_entity *se,
2417 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2418 struct sched_entity *se,
2420 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2421 int force_update) {}
2424 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2426 #ifdef CONFIG_SCHEDSTATS
2427 struct task_struct *tsk = NULL;
2429 if (entity_is_task(se))
2432 if (se->statistics.sleep_start) {
2433 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2438 if (unlikely(delta > se->statistics.sleep_max))
2439 se->statistics.sleep_max = delta;
2441 se->statistics.sleep_start = 0;
2442 se->statistics.sum_sleep_runtime += delta;
2445 account_scheduler_latency(tsk, delta >> 10, 1);
2446 trace_sched_stat_sleep(tsk, delta);
2449 if (se->statistics.block_start) {
2450 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2455 if (unlikely(delta > se->statistics.block_max))
2456 se->statistics.block_max = delta;
2458 se->statistics.block_start = 0;
2459 se->statistics.sum_sleep_runtime += delta;
2462 if (tsk->in_iowait) {
2463 se->statistics.iowait_sum += delta;
2464 se->statistics.iowait_count++;
2465 trace_sched_stat_iowait(tsk, delta);
2468 trace_sched_stat_blocked(tsk, delta);
2471 * Blocking time is in units of nanosecs, so shift by
2472 * 20 to get a milliseconds-range estimation of the
2473 * amount of time that the task spent sleeping:
2475 if (unlikely(prof_on == SLEEP_PROFILING)) {
2476 profile_hits(SLEEP_PROFILING,
2477 (void *)get_wchan(tsk),
2480 account_scheduler_latency(tsk, delta >> 10, 0);
2486 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2488 #ifdef CONFIG_SCHED_DEBUG
2489 s64 d = se->vruntime - cfs_rq->min_vruntime;
2494 if (d > 3*sysctl_sched_latency)
2495 schedstat_inc(cfs_rq, nr_spread_over);
2500 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2502 u64 vruntime = cfs_rq->min_vruntime;
2505 * The 'current' period is already promised to the current tasks,
2506 * however the extra weight of the new task will slow them down a
2507 * little, place the new task so that it fits in the slot that
2508 * stays open at the end.
2510 if (initial && sched_feat(START_DEBIT))
2511 vruntime += sched_vslice(cfs_rq, se);
2513 /* sleeps up to a single latency don't count. */
2515 unsigned long thresh = sysctl_sched_latency;
2518 * Halve their sleep time's effect, to allow
2519 * for a gentler effect of sleepers:
2521 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2527 /* ensure we never gain time by being placed backwards. */
2528 se->vruntime = max_vruntime(se->vruntime, vruntime);
2531 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2534 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2537 * Update the normalized vruntime before updating min_vruntime
2538 * through calling update_curr().
2540 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2541 se->vruntime += cfs_rq->min_vruntime;
2544 * Update run-time statistics of the 'current'.
2546 update_curr(cfs_rq);
2547 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2548 account_entity_enqueue(cfs_rq, se);
2549 update_cfs_shares(cfs_rq);
2551 if (flags & ENQUEUE_WAKEUP) {
2552 place_entity(cfs_rq, se, 0);
2553 enqueue_sleeper(cfs_rq, se);
2556 update_stats_enqueue(cfs_rq, se);
2557 check_spread(cfs_rq, se);
2558 if (se != cfs_rq->curr)
2559 __enqueue_entity(cfs_rq, se);
2562 if (cfs_rq->nr_running == 1) {
2563 list_add_leaf_cfs_rq(cfs_rq);
2564 check_enqueue_throttle(cfs_rq);
2568 static void __clear_buddies_last(struct sched_entity *se)
2570 for_each_sched_entity(se) {
2571 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2572 if (cfs_rq->last == se)
2573 cfs_rq->last = NULL;
2579 static void __clear_buddies_next(struct sched_entity *se)
2581 for_each_sched_entity(se) {
2582 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2583 if (cfs_rq->next == se)
2584 cfs_rq->next = NULL;
2590 static void __clear_buddies_skip(struct sched_entity *se)
2592 for_each_sched_entity(se) {
2593 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2594 if (cfs_rq->skip == se)
2595 cfs_rq->skip = NULL;
2601 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2603 if (cfs_rq->last == se)
2604 __clear_buddies_last(se);
2606 if (cfs_rq->next == se)
2607 __clear_buddies_next(se);
2609 if (cfs_rq->skip == se)
2610 __clear_buddies_skip(se);
2613 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2616 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2619 * Update run-time statistics of the 'current'.
2621 update_curr(cfs_rq);
2622 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2624 update_stats_dequeue(cfs_rq, se);
2625 if (flags & DEQUEUE_SLEEP) {
2626 #ifdef CONFIG_SCHEDSTATS
2627 if (entity_is_task(se)) {
2628 struct task_struct *tsk = task_of(se);
2630 if (tsk->state & TASK_INTERRUPTIBLE)
2631 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2632 if (tsk->state & TASK_UNINTERRUPTIBLE)
2633 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2638 clear_buddies(cfs_rq, se);
2640 if (se != cfs_rq->curr)
2641 __dequeue_entity(cfs_rq, se);
2643 account_entity_dequeue(cfs_rq, se);
2646 * Normalize the entity after updating the min_vruntime because the
2647 * update can refer to the ->curr item and we need to reflect this
2648 * movement in our normalized position.
2650 if (!(flags & DEQUEUE_SLEEP))
2651 se->vruntime -= cfs_rq->min_vruntime;
2653 /* return excess runtime on last dequeue */
2654 return_cfs_rq_runtime(cfs_rq);
2656 update_min_vruntime(cfs_rq);
2657 update_cfs_shares(cfs_rq);
2661 * Preempt the current task with a newly woken task if needed:
2664 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2666 unsigned long ideal_runtime, delta_exec;
2667 struct sched_entity *se;
2670 ideal_runtime = sched_slice(cfs_rq, curr);
2671 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2672 if (delta_exec > ideal_runtime) {
2673 resched_task(rq_of(cfs_rq)->curr);
2675 * The current task ran long enough, ensure it doesn't get
2676 * re-elected due to buddy favours.
2678 clear_buddies(cfs_rq, curr);
2683 * Ensure that a task that missed wakeup preemption by a
2684 * narrow margin doesn't have to wait for a full slice.
2685 * This also mitigates buddy induced latencies under load.
2687 if (delta_exec < sysctl_sched_min_granularity)
2690 se = __pick_first_entity(cfs_rq);
2691 delta = curr->vruntime - se->vruntime;
2696 if (delta > ideal_runtime)
2697 resched_task(rq_of(cfs_rq)->curr);
2701 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2703 /* 'current' is not kept within the tree. */
2706 * Any task has to be enqueued before it get to execute on
2707 * a CPU. So account for the time it spent waiting on the
2710 update_stats_wait_end(cfs_rq, se);
2711 __dequeue_entity(cfs_rq, se);
2714 update_stats_curr_start(cfs_rq, se);
2716 #ifdef CONFIG_SCHEDSTATS
2718 * Track our maximum slice length, if the CPU's load is at
2719 * least twice that of our own weight (i.e. dont track it
2720 * when there are only lesser-weight tasks around):
2722 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2723 se->statistics.slice_max = max(se->statistics.slice_max,
2724 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2727 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2731 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2734 * Pick the next process, keeping these things in mind, in this order:
2735 * 1) keep things fair between processes/task groups
2736 * 2) pick the "next" process, since someone really wants that to run
2737 * 3) pick the "last" process, for cache locality
2738 * 4) do not run the "skip" process, if something else is available
2740 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2742 struct sched_entity *se = __pick_first_entity(cfs_rq);
2743 struct sched_entity *left = se;
2746 * Avoid running the skip buddy, if running something else can
2747 * be done without getting too unfair.
2749 if (cfs_rq->skip == se) {
2750 struct sched_entity *second = __pick_next_entity(se);
2751 if (second && wakeup_preempt_entity(second, left) < 1)
2756 * Prefer last buddy, try to return the CPU to a preempted task.
2758 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2762 * Someone really wants this to run. If it's not unfair, run it.
2764 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2767 clear_buddies(cfs_rq, se);
2772 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2774 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2777 * If still on the runqueue then deactivate_task()
2778 * was not called and update_curr() has to be done:
2781 update_curr(cfs_rq);
2783 /* throttle cfs_rqs exceeding runtime */
2784 check_cfs_rq_runtime(cfs_rq);
2786 check_spread(cfs_rq, prev);
2788 update_stats_wait_start(cfs_rq, prev);
2789 /* Put 'current' back into the tree. */
2790 __enqueue_entity(cfs_rq, prev);
2791 /* in !on_rq case, update occurred at dequeue */
2792 update_entity_load_avg(prev, 1);
2794 cfs_rq->curr = NULL;
2798 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2801 * Update run-time statistics of the 'current'.
2803 update_curr(cfs_rq);
2806 * Ensure that runnable average is periodically updated.
2808 update_entity_load_avg(curr, 1);
2809 update_cfs_rq_blocked_load(cfs_rq, 1);
2810 update_cfs_shares(cfs_rq);
2812 #ifdef CONFIG_SCHED_HRTICK
2814 * queued ticks are scheduled to match the slice, so don't bother
2815 * validating it and just reschedule.
2818 resched_task(rq_of(cfs_rq)->curr);
2822 * don't let the period tick interfere with the hrtick preemption
2824 if (!sched_feat(DOUBLE_TICK) &&
2825 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2829 if (cfs_rq->nr_running > 1)
2830 check_preempt_tick(cfs_rq, curr);
2834 /**************************************************
2835 * CFS bandwidth control machinery
2838 #ifdef CONFIG_CFS_BANDWIDTH
2840 #ifdef HAVE_JUMP_LABEL
2841 static struct static_key __cfs_bandwidth_used;
2843 static inline bool cfs_bandwidth_used(void)
2845 return static_key_false(&__cfs_bandwidth_used);
2848 void cfs_bandwidth_usage_inc(void)
2850 static_key_slow_inc(&__cfs_bandwidth_used);
2853 void cfs_bandwidth_usage_dec(void)
2855 static_key_slow_dec(&__cfs_bandwidth_used);
2857 #else /* HAVE_JUMP_LABEL */
2858 static bool cfs_bandwidth_used(void)
2863 void cfs_bandwidth_usage_inc(void) {}
2864 void cfs_bandwidth_usage_dec(void) {}
2865 #endif /* HAVE_JUMP_LABEL */
2868 * default period for cfs group bandwidth.
2869 * default: 0.1s, units: nanoseconds
2871 static inline u64 default_cfs_period(void)
2873 return 100000000ULL;
2876 static inline u64 sched_cfs_bandwidth_slice(void)
2878 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2882 * Replenish runtime according to assigned quota and update expiration time.
2883 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2884 * additional synchronization around rq->lock.
2886 * requires cfs_b->lock
2888 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2892 if (cfs_b->quota == RUNTIME_INF)
2895 now = sched_clock_cpu(smp_processor_id());
2896 cfs_b->runtime = cfs_b->quota;
2897 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2900 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2902 return &tg->cfs_bandwidth;
2905 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2906 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2908 if (unlikely(cfs_rq->throttle_count))
2909 return cfs_rq->throttled_clock_task;
2911 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2914 /* returns 0 on failure to allocate runtime */
2915 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2917 struct task_group *tg = cfs_rq->tg;
2918 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2919 u64 amount = 0, min_amount, expires;
2921 /* note: this is a positive sum as runtime_remaining <= 0 */
2922 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2924 raw_spin_lock(&cfs_b->lock);
2925 if (cfs_b->quota == RUNTIME_INF)
2926 amount = min_amount;
2929 * If the bandwidth pool has become inactive, then at least one
2930 * period must have elapsed since the last consumption.
2931 * Refresh the global state and ensure bandwidth timer becomes
2934 if (!cfs_b->timer_active) {
2935 __refill_cfs_bandwidth_runtime(cfs_b);
2936 __start_cfs_bandwidth(cfs_b);
2939 if (cfs_b->runtime > 0) {
2940 amount = min(cfs_b->runtime, min_amount);
2941 cfs_b->runtime -= amount;
2945 expires = cfs_b->runtime_expires;
2946 raw_spin_unlock(&cfs_b->lock);
2948 cfs_rq->runtime_remaining += amount;
2950 * we may have advanced our local expiration to account for allowed
2951 * spread between our sched_clock and the one on which runtime was
2954 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2955 cfs_rq->runtime_expires = expires;
2957 return cfs_rq->runtime_remaining > 0;
2961 * Note: This depends on the synchronization provided by sched_clock and the
2962 * fact that rq->clock snapshots this value.
2964 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2966 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2968 /* if the deadline is ahead of our clock, nothing to do */
2969 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2972 if (cfs_rq->runtime_remaining < 0)
2976 * If the local deadline has passed we have to consider the
2977 * possibility that our sched_clock is 'fast' and the global deadline
2978 * has not truly expired.
2980 * Fortunately we can check determine whether this the case by checking
2981 * whether the global deadline has advanced.
2984 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2985 /* extend local deadline, drift is bounded above by 2 ticks */
2986 cfs_rq->runtime_expires += TICK_NSEC;
2988 /* global deadline is ahead, expiration has passed */
2989 cfs_rq->runtime_remaining = 0;
2993 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2994 unsigned long delta_exec)
2996 /* dock delta_exec before expiring quota (as it could span periods) */
2997 cfs_rq->runtime_remaining -= delta_exec;
2998 expire_cfs_rq_runtime(cfs_rq);
3000 if (likely(cfs_rq->runtime_remaining > 0))
3004 * if we're unable to extend our runtime we resched so that the active
3005 * hierarchy can be throttled
3007 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3008 resched_task(rq_of(cfs_rq)->curr);
3011 static __always_inline
3012 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
3014 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3017 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3020 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3022 return cfs_bandwidth_used() && cfs_rq->throttled;
3025 /* check whether cfs_rq, or any parent, is throttled */
3026 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3028 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3032 * Ensure that neither of the group entities corresponding to src_cpu or
3033 * dest_cpu are members of a throttled hierarchy when performing group
3034 * load-balance operations.
3036 static inline int throttled_lb_pair(struct task_group *tg,
3037 int src_cpu, int dest_cpu)
3039 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3041 src_cfs_rq = tg->cfs_rq[src_cpu];
3042 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3044 return throttled_hierarchy(src_cfs_rq) ||
3045 throttled_hierarchy(dest_cfs_rq);
3048 /* updated child weight may affect parent so we have to do this bottom up */
3049 static int tg_unthrottle_up(struct task_group *tg, void *data)
3051 struct rq *rq = data;
3052 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3054 cfs_rq->throttle_count--;
3056 if (!cfs_rq->throttle_count) {
3057 /* adjust cfs_rq_clock_task() */
3058 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3059 cfs_rq->throttled_clock_task;
3066 static int tg_throttle_down(struct task_group *tg, void *data)
3068 struct rq *rq = data;
3069 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3071 /* group is entering throttled state, stop time */
3072 if (!cfs_rq->throttle_count)
3073 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3074 cfs_rq->throttle_count++;
3079 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3081 struct rq *rq = rq_of(cfs_rq);
3082 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3083 struct sched_entity *se;
3084 long task_delta, dequeue = 1;
3086 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3088 /* freeze hierarchy runnable averages while throttled */
3090 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3093 task_delta = cfs_rq->h_nr_running;
3094 for_each_sched_entity(se) {
3095 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3096 /* throttled entity or throttle-on-deactivate */
3101 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3102 qcfs_rq->h_nr_running -= task_delta;
3104 if (qcfs_rq->load.weight)
3109 rq->nr_running -= task_delta;
3111 cfs_rq->throttled = 1;
3112 cfs_rq->throttled_clock = rq_clock(rq);
3113 raw_spin_lock(&cfs_b->lock);
3114 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3115 if (!cfs_b->timer_active)
3116 __start_cfs_bandwidth(cfs_b);
3117 raw_spin_unlock(&cfs_b->lock);
3120 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3122 struct rq *rq = rq_of(cfs_rq);
3123 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3124 struct sched_entity *se;
3128 se = cfs_rq->tg->se[cpu_of(rq)];
3130 cfs_rq->throttled = 0;
3132 update_rq_clock(rq);
3134 raw_spin_lock(&cfs_b->lock);
3135 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3136 list_del_rcu(&cfs_rq->throttled_list);
3137 raw_spin_unlock(&cfs_b->lock);
3139 /* update hierarchical throttle state */
3140 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3142 if (!cfs_rq->load.weight)
3145 task_delta = cfs_rq->h_nr_running;
3146 for_each_sched_entity(se) {
3150 cfs_rq = cfs_rq_of(se);
3152 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3153 cfs_rq->h_nr_running += task_delta;
3155 if (cfs_rq_throttled(cfs_rq))
3160 rq->nr_running += task_delta;
3162 /* determine whether we need to wake up potentially idle cpu */
3163 if (rq->curr == rq->idle && rq->cfs.nr_running)
3164 resched_task(rq->curr);
3167 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3168 u64 remaining, u64 expires)
3170 struct cfs_rq *cfs_rq;
3171 u64 runtime = remaining;
3174 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3176 struct rq *rq = rq_of(cfs_rq);
3178 raw_spin_lock(&rq->lock);
3179 if (!cfs_rq_throttled(cfs_rq))
3182 runtime = -cfs_rq->runtime_remaining + 1;
3183 if (runtime > remaining)
3184 runtime = remaining;
3185 remaining -= runtime;
3187 cfs_rq->runtime_remaining += runtime;
3188 cfs_rq->runtime_expires = expires;
3190 /* we check whether we're throttled above */
3191 if (cfs_rq->runtime_remaining > 0)
3192 unthrottle_cfs_rq(cfs_rq);
3195 raw_spin_unlock(&rq->lock);
3206 * Responsible for refilling a task_group's bandwidth and unthrottling its
3207 * cfs_rqs as appropriate. If there has been no activity within the last
3208 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3209 * used to track this state.
3211 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3213 u64 runtime, runtime_expires;
3214 int idle = 1, throttled;
3216 raw_spin_lock(&cfs_b->lock);
3217 /* no need to continue the timer with no bandwidth constraint */
3218 if (cfs_b->quota == RUNTIME_INF)
3221 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3222 /* idle depends on !throttled (for the case of a large deficit) */
3223 idle = cfs_b->idle && !throttled;
3224 cfs_b->nr_periods += overrun;
3226 /* if we're going inactive then everything else can be deferred */
3231 * if we have relooped after returning idle once, we need to update our
3232 * status as actually running, so that other cpus doing
3233 * __start_cfs_bandwidth will stop trying to cancel us.
3235 cfs_b->timer_active = 1;
3237 __refill_cfs_bandwidth_runtime(cfs_b);
3240 /* mark as potentially idle for the upcoming period */
3245 /* account preceding periods in which throttling occurred */
3246 cfs_b->nr_throttled += overrun;
3249 * There are throttled entities so we must first use the new bandwidth
3250 * to unthrottle them before making it generally available. This
3251 * ensures that all existing debts will be paid before a new cfs_rq is
3254 runtime = cfs_b->runtime;
3255 runtime_expires = cfs_b->runtime_expires;
3259 * This check is repeated as we are holding onto the new bandwidth
3260 * while we unthrottle. This can potentially race with an unthrottled
3261 * group trying to acquire new bandwidth from the global pool.
3263 while (throttled && runtime > 0) {
3264 raw_spin_unlock(&cfs_b->lock);
3265 /* we can't nest cfs_b->lock while distributing bandwidth */
3266 runtime = distribute_cfs_runtime(cfs_b, runtime,
3268 raw_spin_lock(&cfs_b->lock);
3270 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3273 /* return (any) remaining runtime */
3274 cfs_b->runtime = runtime;
3276 * While we are ensured activity in the period following an
3277 * unthrottle, this also covers the case in which the new bandwidth is
3278 * insufficient to cover the existing bandwidth deficit. (Forcing the
3279 * timer to remain active while there are any throttled entities.)
3284 cfs_b->timer_active = 0;
3285 raw_spin_unlock(&cfs_b->lock);
3290 /* a cfs_rq won't donate quota below this amount */
3291 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3292 /* minimum remaining period time to redistribute slack quota */
3293 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3294 /* how long we wait to gather additional slack before distributing */
3295 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3298 * Are we near the end of the current quota period?
3300 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3301 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3302 * migrate_hrtimers, base is never cleared, so we are fine.
3304 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3306 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3309 /* if the call-back is running a quota refresh is already occurring */
3310 if (hrtimer_callback_running(refresh_timer))
3313 /* is a quota refresh about to occur? */
3314 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3315 if (remaining < min_expire)
3321 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3323 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3325 /* if there's a quota refresh soon don't bother with slack */
3326 if (runtime_refresh_within(cfs_b, min_left))
3329 start_bandwidth_timer(&cfs_b->slack_timer,
3330 ns_to_ktime(cfs_bandwidth_slack_period));
3333 /* we know any runtime found here is valid as update_curr() precedes return */
3334 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3336 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3337 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3339 if (slack_runtime <= 0)
3342 raw_spin_lock(&cfs_b->lock);
3343 if (cfs_b->quota != RUNTIME_INF &&
3344 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3345 cfs_b->runtime += slack_runtime;
3347 /* we are under rq->lock, defer unthrottling using a timer */
3348 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3349 !list_empty(&cfs_b->throttled_cfs_rq))
3350 start_cfs_slack_bandwidth(cfs_b);
3352 raw_spin_unlock(&cfs_b->lock);
3354 /* even if it's not valid for return we don't want to try again */
3355 cfs_rq->runtime_remaining -= slack_runtime;
3358 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3360 if (!cfs_bandwidth_used())
3363 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3366 __return_cfs_rq_runtime(cfs_rq);
3370 * This is done with a timer (instead of inline with bandwidth return) since
3371 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3373 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3375 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3378 /* confirm we're still not at a refresh boundary */
3379 raw_spin_lock(&cfs_b->lock);
3380 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3381 raw_spin_unlock(&cfs_b->lock);
3385 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3386 runtime = cfs_b->runtime;
3389 expires = cfs_b->runtime_expires;
3390 raw_spin_unlock(&cfs_b->lock);
3395 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3397 raw_spin_lock(&cfs_b->lock);
3398 if (expires == cfs_b->runtime_expires)
3399 cfs_b->runtime = runtime;
3400 raw_spin_unlock(&cfs_b->lock);
3404 * When a group wakes up we want to make sure that its quota is not already
3405 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3406 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3408 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3410 if (!cfs_bandwidth_used())
3413 /* an active group must be handled by the update_curr()->put() path */
3414 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3417 /* ensure the group is not already throttled */
3418 if (cfs_rq_throttled(cfs_rq))
3421 /* update runtime allocation */
3422 account_cfs_rq_runtime(cfs_rq, 0);
3423 if (cfs_rq->runtime_remaining <= 0)
3424 throttle_cfs_rq(cfs_rq);
3427 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3428 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3430 if (!cfs_bandwidth_used())
3433 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3437 * it's possible for a throttled entity to be forced into a running
3438 * state (e.g. set_curr_task), in this case we're finished.
3440 if (cfs_rq_throttled(cfs_rq))
3443 throttle_cfs_rq(cfs_rq);
3446 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3448 struct cfs_bandwidth *cfs_b =
3449 container_of(timer, struct cfs_bandwidth, slack_timer);
3450 do_sched_cfs_slack_timer(cfs_b);
3452 return HRTIMER_NORESTART;
3455 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3457 struct cfs_bandwidth *cfs_b =
3458 container_of(timer, struct cfs_bandwidth, period_timer);
3464 now = hrtimer_cb_get_time(timer);
3465 overrun = hrtimer_forward(timer, now, cfs_b->period);
3470 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3473 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3476 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3478 raw_spin_lock_init(&cfs_b->lock);
3480 cfs_b->quota = RUNTIME_INF;
3481 cfs_b->period = ns_to_ktime(default_cfs_period());
3483 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3484 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3485 cfs_b->period_timer.function = sched_cfs_period_timer;
3486 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3487 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3490 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3492 cfs_rq->runtime_enabled = 0;
3493 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3496 /* requires cfs_b->lock, may release to reprogram timer */
3497 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3500 * The timer may be active because we're trying to set a new bandwidth
3501 * period or because we're racing with the tear-down path
3502 * (timer_active==0 becomes visible before the hrtimer call-back
3503 * terminates). In either case we ensure that it's re-programmed
3505 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3506 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3507 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3508 raw_spin_unlock(&cfs_b->lock);
3510 raw_spin_lock(&cfs_b->lock);
3511 /* if someone else restarted the timer then we're done */
3512 if (cfs_b->timer_active)
3516 cfs_b->timer_active = 1;
3517 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3520 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3522 hrtimer_cancel(&cfs_b->period_timer);
3523 hrtimer_cancel(&cfs_b->slack_timer);
3526 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3528 struct cfs_rq *cfs_rq;
3530 for_each_leaf_cfs_rq(rq, cfs_rq) {
3531 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3533 if (!cfs_rq->runtime_enabled)
3537 * clock_task is not advancing so we just need to make sure
3538 * there's some valid quota amount
3540 cfs_rq->runtime_remaining = cfs_b->quota;
3541 if (cfs_rq_throttled(cfs_rq))
3542 unthrottle_cfs_rq(cfs_rq);
3546 #else /* CONFIG_CFS_BANDWIDTH */
3547 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3549 return rq_clock_task(rq_of(cfs_rq));
3552 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3553 unsigned long delta_exec) {}
3554 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3555 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3556 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3558 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3563 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3568 static inline int throttled_lb_pair(struct task_group *tg,
3569 int src_cpu, int dest_cpu)
3574 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3576 #ifdef CONFIG_FAIR_GROUP_SCHED
3577 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3580 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3584 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3585 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3587 #endif /* CONFIG_CFS_BANDWIDTH */
3589 /**************************************************
3590 * CFS operations on tasks:
3593 #ifdef CONFIG_SCHED_HRTICK
3594 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3596 struct sched_entity *se = &p->se;
3597 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3599 WARN_ON(task_rq(p) != rq);
3601 if (cfs_rq->nr_running > 1) {
3602 u64 slice = sched_slice(cfs_rq, se);
3603 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3604 s64 delta = slice - ran;
3613 * Don't schedule slices shorter than 10000ns, that just
3614 * doesn't make sense. Rely on vruntime for fairness.
3617 delta = max_t(s64, 10000LL, delta);
3619 hrtick_start(rq, delta);
3624 * called from enqueue/dequeue and updates the hrtick when the
3625 * current task is from our class and nr_running is low enough
3628 static void hrtick_update(struct rq *rq)
3630 struct task_struct *curr = rq->curr;
3632 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3635 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3636 hrtick_start_fair(rq, curr);
3638 #else /* !CONFIG_SCHED_HRTICK */
3640 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3644 static inline void hrtick_update(struct rq *rq)
3650 * The enqueue_task method is called before nr_running is
3651 * increased. Here we update the fair scheduling stats and
3652 * then put the task into the rbtree:
3655 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3657 struct cfs_rq *cfs_rq;
3658 struct sched_entity *se = &p->se;
3660 for_each_sched_entity(se) {
3663 cfs_rq = cfs_rq_of(se);
3664 enqueue_entity(cfs_rq, se, flags);
3667 * end evaluation on encountering a throttled cfs_rq
3669 * note: in the case of encountering a throttled cfs_rq we will
3670 * post the final h_nr_running increment below.
3672 if (cfs_rq_throttled(cfs_rq))
3674 cfs_rq->h_nr_running++;
3676 flags = ENQUEUE_WAKEUP;
3679 for_each_sched_entity(se) {
3680 cfs_rq = cfs_rq_of(se);
3681 cfs_rq->h_nr_running++;
3683 if (cfs_rq_throttled(cfs_rq))
3686 update_cfs_shares(cfs_rq);
3687 update_entity_load_avg(se, 1);
3691 update_rq_runnable_avg(rq, rq->nr_running);
3697 static void set_next_buddy(struct sched_entity *se);
3700 * The dequeue_task method is called before nr_running is
3701 * decreased. We remove the task from the rbtree and
3702 * update the fair scheduling stats:
3704 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3706 struct cfs_rq *cfs_rq;
3707 struct sched_entity *se = &p->se;
3708 int task_sleep = flags & DEQUEUE_SLEEP;
3710 for_each_sched_entity(se) {
3711 cfs_rq = cfs_rq_of(se);
3712 dequeue_entity(cfs_rq, se, flags);
3715 * end evaluation on encountering a throttled cfs_rq
3717 * note: in the case of encountering a throttled cfs_rq we will
3718 * post the final h_nr_running decrement below.
3720 if (cfs_rq_throttled(cfs_rq))
3722 cfs_rq->h_nr_running--;
3724 /* Don't dequeue parent if it has other entities besides us */
3725 if (cfs_rq->load.weight) {
3727 * Bias pick_next to pick a task from this cfs_rq, as
3728 * p is sleeping when it is within its sched_slice.
3730 if (task_sleep && parent_entity(se))
3731 set_next_buddy(parent_entity(se));
3733 /* avoid re-evaluating load for this entity */
3734 se = parent_entity(se);
3737 flags |= DEQUEUE_SLEEP;
3740 for_each_sched_entity(se) {
3741 cfs_rq = cfs_rq_of(se);
3742 cfs_rq->h_nr_running--;
3744 if (cfs_rq_throttled(cfs_rq))
3747 update_cfs_shares(cfs_rq);
3748 update_entity_load_avg(se, 1);
3753 update_rq_runnable_avg(rq, 1);
3759 /* Used instead of source_load when we know the type == 0 */
3760 static unsigned long weighted_cpuload(const int cpu)
3762 return cpu_rq(cpu)->cfs.runnable_load_avg;
3766 * Return a low guess at the load of a migration-source cpu weighted
3767 * according to the scheduling class and "nice" value.
3769 * We want to under-estimate the load of migration sources, to
3770 * balance conservatively.
3772 static unsigned long source_load(int cpu, int type)
3774 struct rq *rq = cpu_rq(cpu);
3775 unsigned long total = weighted_cpuload(cpu);
3777 if (type == 0 || !sched_feat(LB_BIAS))
3780 return min(rq->cpu_load[type-1], total);
3784 * Return a high guess at the load of a migration-target cpu weighted
3785 * according to the scheduling class and "nice" value.
3787 static unsigned long target_load(int cpu, int type)
3789 struct rq *rq = cpu_rq(cpu);
3790 unsigned long total = weighted_cpuload(cpu);
3792 if (type == 0 || !sched_feat(LB_BIAS))
3795 return max(rq->cpu_load[type-1], total);
3798 static unsigned long power_of(int cpu)
3800 return cpu_rq(cpu)->cpu_power;
3803 static unsigned long cpu_avg_load_per_task(int cpu)
3805 struct rq *rq = cpu_rq(cpu);
3806 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3807 unsigned long load_avg = rq->cfs.runnable_load_avg;
3810 return load_avg / nr_running;
3815 static void record_wakee(struct task_struct *p)
3818 * Rough decay (wiping) for cost saving, don't worry
3819 * about the boundary, really active task won't care
3822 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3823 current->wakee_flips = 0;
3824 current->wakee_flip_decay_ts = jiffies;
3827 if (current->last_wakee != p) {
3828 current->last_wakee = p;
3829 current->wakee_flips++;
3833 static void task_waking_fair(struct task_struct *p)
3835 struct sched_entity *se = &p->se;
3836 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3839 #ifndef CONFIG_64BIT
3840 u64 min_vruntime_copy;
3843 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3845 min_vruntime = cfs_rq->min_vruntime;
3846 } while (min_vruntime != min_vruntime_copy);
3848 min_vruntime = cfs_rq->min_vruntime;
3851 se->vruntime -= min_vruntime;
3855 #ifdef CONFIG_FAIR_GROUP_SCHED
3857 * effective_load() calculates the load change as seen from the root_task_group
3859 * Adding load to a group doesn't make a group heavier, but can cause movement
3860 * of group shares between cpus. Assuming the shares were perfectly aligned one
3861 * can calculate the shift in shares.
3863 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3864 * on this @cpu and results in a total addition (subtraction) of @wg to the
3865 * total group weight.
3867 * Given a runqueue weight distribution (rw_i) we can compute a shares
3868 * distribution (s_i) using:
3870 * s_i = rw_i / \Sum rw_j (1)
3872 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3873 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3874 * shares distribution (s_i):
3876 * rw_i = { 2, 4, 1, 0 }
3877 * s_i = { 2/7, 4/7, 1/7, 0 }
3879 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3880 * task used to run on and the CPU the waker is running on), we need to
3881 * compute the effect of waking a task on either CPU and, in case of a sync
3882 * wakeup, compute the effect of the current task going to sleep.
3884 * So for a change of @wl to the local @cpu with an overall group weight change
3885 * of @wl we can compute the new shares distribution (s'_i) using:
3887 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3889 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3890 * differences in waking a task to CPU 0. The additional task changes the
3891 * weight and shares distributions like:
3893 * rw'_i = { 3, 4, 1, 0 }
3894 * s'_i = { 3/8, 4/8, 1/8, 0 }
3896 * We can then compute the difference in effective weight by using:
3898 * dw_i = S * (s'_i - s_i) (3)
3900 * Where 'S' is the group weight as seen by its parent.
3902 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3903 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3904 * 4/7) times the weight of the group.
3906 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3908 struct sched_entity *se = tg->se[cpu];
3910 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3913 for_each_sched_entity(se) {
3919 * W = @wg + \Sum rw_j
3921 W = wg + calc_tg_weight(tg, se->my_q);
3926 w = se->my_q->load.weight + wl;
3929 * wl = S * s'_i; see (2)
3932 wl = (w * tg->shares) / W;
3937 * Per the above, wl is the new se->load.weight value; since
3938 * those are clipped to [MIN_SHARES, ...) do so now. See
3939 * calc_cfs_shares().
3941 if (wl < MIN_SHARES)
3945 * wl = dw_i = S * (s'_i - s_i); see (3)
3947 wl -= se->load.weight;
3950 * Recursively apply this logic to all parent groups to compute
3951 * the final effective load change on the root group. Since
3952 * only the @tg group gets extra weight, all parent groups can
3953 * only redistribute existing shares. @wl is the shift in shares
3954 * resulting from this level per the above.
3963 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3970 static int wake_wide(struct task_struct *p)
3972 int factor = this_cpu_read(sd_llc_size);
3975 * Yeah, it's the switching-frequency, could means many wakee or
3976 * rapidly switch, use factor here will just help to automatically
3977 * adjust the loose-degree, so bigger node will lead to more pull.
3979 if (p->wakee_flips > factor) {
3981 * wakee is somewhat hot, it needs certain amount of cpu
3982 * resource, so if waker is far more hot, prefer to leave
3985 if (current->wakee_flips > (factor * p->wakee_flips))
3992 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3994 s64 this_load, load;
3995 int idx, this_cpu, prev_cpu;
3996 unsigned long tl_per_task;
3997 struct task_group *tg;
3998 unsigned long weight;
4002 * If we wake multiple tasks be careful to not bounce
4003 * ourselves around too much.
4009 this_cpu = smp_processor_id();
4010 prev_cpu = task_cpu(p);
4011 load = source_load(prev_cpu, idx);
4012 this_load = target_load(this_cpu, idx);
4015 * If sync wakeup then subtract the (maximum possible)
4016 * effect of the currently running task from the load
4017 * of the current CPU:
4020 tg = task_group(current);
4021 weight = current->se.load.weight;
4023 this_load += effective_load(tg, this_cpu, -weight, -weight);
4024 load += effective_load(tg, prev_cpu, 0, -weight);
4028 weight = p->se.load.weight;
4031 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4032 * due to the sync cause above having dropped this_load to 0, we'll
4033 * always have an imbalance, but there's really nothing you can do
4034 * about that, so that's good too.
4036 * Otherwise check if either cpus are near enough in load to allow this
4037 * task to be woken on this_cpu.
4039 if (this_load > 0) {
4040 s64 this_eff_load, prev_eff_load;
4042 this_eff_load = 100;
4043 this_eff_load *= power_of(prev_cpu);
4044 this_eff_load *= this_load +
4045 effective_load(tg, this_cpu, weight, weight);
4047 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4048 prev_eff_load *= power_of(this_cpu);
4049 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4051 balanced = this_eff_load <= prev_eff_load;
4056 * If the currently running task will sleep within
4057 * a reasonable amount of time then attract this newly
4060 if (sync && balanced)
4063 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4064 tl_per_task = cpu_avg_load_per_task(this_cpu);
4067 (this_load <= load &&
4068 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4070 * This domain has SD_WAKE_AFFINE and
4071 * p is cache cold in this domain, and
4072 * there is no bad imbalance.
4074 schedstat_inc(sd, ttwu_move_affine);
4075 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4083 * find_idlest_group finds and returns the least busy CPU group within the
4086 static struct sched_group *
4087 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4088 int this_cpu, int load_idx)
4090 struct sched_group *idlest = NULL, *group = sd->groups;
4091 unsigned long min_load = ULONG_MAX, this_load = 0;
4092 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4095 unsigned long load, avg_load;
4099 /* Skip over this group if it has no CPUs allowed */
4100 if (!cpumask_intersects(sched_group_cpus(group),
4101 tsk_cpus_allowed(p)))
4104 local_group = cpumask_test_cpu(this_cpu,
4105 sched_group_cpus(group));
4107 /* Tally up the load of all CPUs in the group */
4110 for_each_cpu(i, sched_group_cpus(group)) {
4111 /* Bias balancing toward cpus of our domain */
4113 load = source_load(i, load_idx);
4115 load = target_load(i, load_idx);
4120 /* Adjust by relative CPU power of the group */
4121 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4124 this_load = avg_load;
4125 } else if (avg_load < min_load) {
4126 min_load = avg_load;
4129 } while (group = group->next, group != sd->groups);
4131 if (!idlest || 100*this_load < imbalance*min_load)
4137 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4140 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4142 unsigned long load, min_load = ULONG_MAX;
4146 /* Traverse only the allowed CPUs */
4147 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4148 load = weighted_cpuload(i);
4150 if (load < min_load || (load == min_load && i == this_cpu)) {
4160 * Try and locate an idle CPU in the sched_domain.
4162 static int select_idle_sibling(struct task_struct *p, int target)
4164 struct sched_domain *sd;
4165 struct sched_group *sg;
4166 int i = task_cpu(p);
4168 if (idle_cpu(target))
4172 * If the prevous cpu is cache affine and idle, don't be stupid.
4174 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4178 * Otherwise, iterate the domains and find an elegible idle cpu.
4180 sd = rcu_dereference(per_cpu(sd_llc, target));
4181 for_each_lower_domain(sd) {
4184 if (!cpumask_intersects(sched_group_cpus(sg),
4185 tsk_cpus_allowed(p)))
4188 for_each_cpu(i, sched_group_cpus(sg)) {
4189 if (i == target || !idle_cpu(i))
4193 target = cpumask_first_and(sched_group_cpus(sg),
4194 tsk_cpus_allowed(p));
4198 } while (sg != sd->groups);
4205 * sched_balance_self: balance the current task (running on cpu) in domains
4206 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4209 * Balance, ie. select the least loaded group.
4211 * Returns the target CPU number, or the same CPU if no balancing is needed.
4213 * preempt must be disabled.
4216 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4218 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4219 int cpu = smp_processor_id();
4221 int want_affine = 0;
4222 int sync = wake_flags & WF_SYNC;
4224 if (p->nr_cpus_allowed == 1)
4227 if (sd_flag & SD_BALANCE_WAKE) {
4228 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4234 for_each_domain(cpu, tmp) {
4235 if (!(tmp->flags & SD_LOAD_BALANCE))
4239 * If both cpu and prev_cpu are part of this domain,
4240 * cpu is a valid SD_WAKE_AFFINE target.
4242 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4243 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4248 if (tmp->flags & sd_flag)
4253 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4256 new_cpu = select_idle_sibling(p, prev_cpu);
4261 int load_idx = sd->forkexec_idx;
4262 struct sched_group *group;
4265 if (!(sd->flags & sd_flag)) {
4270 if (sd_flag & SD_BALANCE_WAKE)
4271 load_idx = sd->wake_idx;
4273 group = find_idlest_group(sd, p, cpu, load_idx);
4279 new_cpu = find_idlest_cpu(group, p, cpu);
4280 if (new_cpu == -1 || new_cpu == cpu) {
4281 /* Now try balancing at a lower domain level of cpu */
4286 /* Now try balancing at a lower domain level of new_cpu */
4288 weight = sd->span_weight;
4290 for_each_domain(cpu, tmp) {
4291 if (weight <= tmp->span_weight)
4293 if (tmp->flags & sd_flag)
4296 /* while loop will break here if sd == NULL */
4305 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4306 * cfs_rq_of(p) references at time of call are still valid and identify the
4307 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4308 * other assumptions, including the state of rq->lock, should be made.
4311 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4313 struct sched_entity *se = &p->se;
4314 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4317 * Load tracking: accumulate removed load so that it can be processed
4318 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4319 * to blocked load iff they have a positive decay-count. It can never
4320 * be negative here since on-rq tasks have decay-count == 0.
4322 if (se->avg.decay_count) {
4323 se->avg.decay_count = -__synchronize_entity_decay(se);
4324 atomic_long_add(se->avg.load_avg_contrib,
4325 &cfs_rq->removed_load);
4328 #endif /* CONFIG_SMP */
4330 static unsigned long
4331 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4333 unsigned long gran = sysctl_sched_wakeup_granularity;
4336 * Since its curr running now, convert the gran from real-time
4337 * to virtual-time in his units.
4339 * By using 'se' instead of 'curr' we penalize light tasks, so
4340 * they get preempted easier. That is, if 'se' < 'curr' then
4341 * the resulting gran will be larger, therefore penalizing the
4342 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4343 * be smaller, again penalizing the lighter task.
4345 * This is especially important for buddies when the leftmost
4346 * task is higher priority than the buddy.
4348 return calc_delta_fair(gran, se);
4352 * Should 'se' preempt 'curr'.
4366 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4368 s64 gran, vdiff = curr->vruntime - se->vruntime;
4373 gran = wakeup_gran(curr, se);
4380 static void set_last_buddy(struct sched_entity *se)
4382 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4385 for_each_sched_entity(se)
4386 cfs_rq_of(se)->last = se;
4389 static void set_next_buddy(struct sched_entity *se)
4391 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4394 for_each_sched_entity(se)
4395 cfs_rq_of(se)->next = se;
4398 static void set_skip_buddy(struct sched_entity *se)
4400 for_each_sched_entity(se)
4401 cfs_rq_of(se)->skip = se;
4405 * Preempt the current task with a newly woken task if needed:
4407 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4409 struct task_struct *curr = rq->curr;
4410 struct sched_entity *se = &curr->se, *pse = &p->se;
4411 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4412 int scale = cfs_rq->nr_running >= sched_nr_latency;
4413 int next_buddy_marked = 0;
4415 if (unlikely(se == pse))
4419 * This is possible from callers such as move_task(), in which we
4420 * unconditionally check_prempt_curr() after an enqueue (which may have
4421 * lead to a throttle). This both saves work and prevents false
4422 * next-buddy nomination below.
4424 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4427 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4428 set_next_buddy(pse);
4429 next_buddy_marked = 1;
4433 * We can come here with TIF_NEED_RESCHED already set from new task
4436 * Note: this also catches the edge-case of curr being in a throttled
4437 * group (e.g. via set_curr_task), since update_curr() (in the
4438 * enqueue of curr) will have resulted in resched being set. This
4439 * prevents us from potentially nominating it as a false LAST_BUDDY
4442 if (test_tsk_need_resched(curr))
4445 /* Idle tasks are by definition preempted by non-idle tasks. */
4446 if (unlikely(curr->policy == SCHED_IDLE) &&
4447 likely(p->policy != SCHED_IDLE))
4451 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4452 * is driven by the tick):
4454 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4457 find_matching_se(&se, &pse);
4458 update_curr(cfs_rq_of(se));
4460 if (wakeup_preempt_entity(se, pse) == 1) {
4462 * Bias pick_next to pick the sched entity that is
4463 * triggering this preemption.
4465 if (!next_buddy_marked)
4466 set_next_buddy(pse);
4475 * Only set the backward buddy when the current task is still
4476 * on the rq. This can happen when a wakeup gets interleaved
4477 * with schedule on the ->pre_schedule() or idle_balance()
4478 * point, either of which can * drop the rq lock.
4480 * Also, during early boot the idle thread is in the fair class,
4481 * for obvious reasons its a bad idea to schedule back to it.
4483 if (unlikely(!se->on_rq || curr == rq->idle))
4486 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4490 static struct task_struct *pick_next_task_fair(struct rq *rq)
4492 struct task_struct *p;
4493 struct cfs_rq *cfs_rq = &rq->cfs;
4494 struct sched_entity *se;
4496 if (!cfs_rq->nr_running)
4500 se = pick_next_entity(cfs_rq);
4501 set_next_entity(cfs_rq, se);
4502 cfs_rq = group_cfs_rq(se);
4506 if (hrtick_enabled(rq))
4507 hrtick_start_fair(rq, p);
4513 * Account for a descheduled task:
4515 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4517 struct sched_entity *se = &prev->se;
4518 struct cfs_rq *cfs_rq;
4520 for_each_sched_entity(se) {
4521 cfs_rq = cfs_rq_of(se);
4522 put_prev_entity(cfs_rq, se);
4527 * sched_yield() is very simple
4529 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4531 static void yield_task_fair(struct rq *rq)
4533 struct task_struct *curr = rq->curr;
4534 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4535 struct sched_entity *se = &curr->se;
4538 * Are we the only task in the tree?
4540 if (unlikely(rq->nr_running == 1))
4543 clear_buddies(cfs_rq, se);
4545 if (curr->policy != SCHED_BATCH) {
4546 update_rq_clock(rq);
4548 * Update run-time statistics of the 'current'.
4550 update_curr(cfs_rq);
4552 * Tell update_rq_clock() that we've just updated,
4553 * so we don't do microscopic update in schedule()
4554 * and double the fastpath cost.
4556 rq->skip_clock_update = 1;
4562 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4564 struct sched_entity *se = &p->se;
4566 /* throttled hierarchies are not runnable */
4567 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4570 /* Tell the scheduler that we'd really like pse to run next. */
4573 yield_task_fair(rq);
4579 /**************************************************
4580 * Fair scheduling class load-balancing methods.
4584 * The purpose of load-balancing is to achieve the same basic fairness the
4585 * per-cpu scheduler provides, namely provide a proportional amount of compute
4586 * time to each task. This is expressed in the following equation:
4588 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4590 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4591 * W_i,0 is defined as:
4593 * W_i,0 = \Sum_j w_i,j (2)
4595 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4596 * is derived from the nice value as per prio_to_weight[].
4598 * The weight average is an exponential decay average of the instantaneous
4601 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4603 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4604 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4605 * can also include other factors [XXX].
4607 * To achieve this balance we define a measure of imbalance which follows
4608 * directly from (1):
4610 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4612 * We them move tasks around to minimize the imbalance. In the continuous
4613 * function space it is obvious this converges, in the discrete case we get
4614 * a few fun cases generally called infeasible weight scenarios.
4617 * - infeasible weights;
4618 * - local vs global optima in the discrete case. ]
4623 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4624 * for all i,j solution, we create a tree of cpus that follows the hardware
4625 * topology where each level pairs two lower groups (or better). This results
4626 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4627 * tree to only the first of the previous level and we decrease the frequency
4628 * of load-balance at each level inv. proportional to the number of cpus in
4634 * \Sum { --- * --- * 2^i } = O(n) (5)
4636 * `- size of each group
4637 * | | `- number of cpus doing load-balance
4639 * `- sum over all levels
4641 * Coupled with a limit on how many tasks we can migrate every balance pass,
4642 * this makes (5) the runtime complexity of the balancer.
4644 * An important property here is that each CPU is still (indirectly) connected
4645 * to every other cpu in at most O(log n) steps:
4647 * The adjacency matrix of the resulting graph is given by:
4650 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4653 * And you'll find that:
4655 * A^(log_2 n)_i,j != 0 for all i,j (7)
4657 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4658 * The task movement gives a factor of O(m), giving a convergence complexity
4661 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4666 * In order to avoid CPUs going idle while there's still work to do, new idle
4667 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4668 * tree itself instead of relying on other CPUs to bring it work.
4670 * This adds some complexity to both (5) and (8) but it reduces the total idle
4678 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4681 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4686 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4688 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4690 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4693 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4694 * rewrite all of this once again.]
4697 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4699 enum fbq_type { regular, remote, all };
4701 #define LBF_ALL_PINNED 0x01
4702 #define LBF_NEED_BREAK 0x02
4703 #define LBF_DST_PINNED 0x04
4704 #define LBF_SOME_PINNED 0x08
4707 struct sched_domain *sd;
4715 struct cpumask *dst_grpmask;
4717 enum cpu_idle_type idle;
4719 /* The set of CPUs under consideration for load-balancing */
4720 struct cpumask *cpus;
4725 unsigned int loop_break;
4726 unsigned int loop_max;
4728 enum fbq_type fbq_type;
4732 * move_task - move a task from one runqueue to another runqueue.
4733 * Both runqueues must be locked.
4735 static void move_task(struct task_struct *p, struct lb_env *env)
4737 deactivate_task(env->src_rq, p, 0);
4738 set_task_cpu(p, env->dst_cpu);
4739 activate_task(env->dst_rq, p, 0);
4740 check_preempt_curr(env->dst_rq, p, 0);
4744 * Is this task likely cache-hot:
4747 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4751 if (p->sched_class != &fair_sched_class)
4754 if (unlikely(p->policy == SCHED_IDLE))
4758 * Buddy candidates are cache hot:
4760 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4761 (&p->se == cfs_rq_of(&p->se)->next ||
4762 &p->se == cfs_rq_of(&p->se)->last))
4765 if (sysctl_sched_migration_cost == -1)
4767 if (sysctl_sched_migration_cost == 0)
4770 delta = now - p->se.exec_start;
4772 return delta < (s64)sysctl_sched_migration_cost;
4775 #ifdef CONFIG_NUMA_BALANCING
4776 /* Returns true if the destination node has incurred more faults */
4777 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4779 int src_nid, dst_nid;
4781 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4782 !(env->sd->flags & SD_NUMA)) {
4786 src_nid = cpu_to_node(env->src_cpu);
4787 dst_nid = cpu_to_node(env->dst_cpu);
4789 if (src_nid == dst_nid)
4792 /* Always encourage migration to the preferred node. */
4793 if (dst_nid == p->numa_preferred_nid)
4796 /* If both task and group weight improve, this move is a winner. */
4797 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4798 group_weight(p, dst_nid) > group_weight(p, src_nid))
4805 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4807 int src_nid, dst_nid;
4809 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4812 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4815 src_nid = cpu_to_node(env->src_cpu);
4816 dst_nid = cpu_to_node(env->dst_cpu);
4818 if (src_nid == dst_nid)
4821 /* Migrating away from the preferred node is always bad. */
4822 if (src_nid == p->numa_preferred_nid)
4825 /* If either task or group weight get worse, don't do it. */
4826 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4827 group_weight(p, dst_nid) < group_weight(p, src_nid))
4834 static inline bool migrate_improves_locality(struct task_struct *p,
4840 static inline bool migrate_degrades_locality(struct task_struct *p,
4848 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4851 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4853 int tsk_cache_hot = 0;
4855 * We do not migrate tasks that are:
4856 * 1) throttled_lb_pair, or
4857 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4858 * 3) running (obviously), or
4859 * 4) are cache-hot on their current CPU.
4861 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4864 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4867 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4869 env->flags |= LBF_SOME_PINNED;
4872 * Remember if this task can be migrated to any other cpu in
4873 * our sched_group. We may want to revisit it if we couldn't
4874 * meet load balance goals by pulling other tasks on src_cpu.
4876 * Also avoid computing new_dst_cpu if we have already computed
4877 * one in current iteration.
4879 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4882 /* Prevent to re-select dst_cpu via env's cpus */
4883 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4884 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4885 env->flags |= LBF_DST_PINNED;
4886 env->new_dst_cpu = cpu;
4894 /* Record that we found atleast one task that could run on dst_cpu */
4895 env->flags &= ~LBF_ALL_PINNED;
4897 if (task_running(env->src_rq, p)) {
4898 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4903 * Aggressive migration if:
4904 * 1) destination numa is preferred
4905 * 2) task is cache cold, or
4906 * 3) too many balance attempts have failed.
4908 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4910 tsk_cache_hot = migrate_degrades_locality(p, env);
4912 if (migrate_improves_locality(p, env)) {
4913 #ifdef CONFIG_SCHEDSTATS
4914 if (tsk_cache_hot) {
4915 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4916 schedstat_inc(p, se.statistics.nr_forced_migrations);
4922 if (!tsk_cache_hot ||
4923 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4925 if (tsk_cache_hot) {
4926 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4927 schedstat_inc(p, se.statistics.nr_forced_migrations);
4933 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4938 * move_one_task tries to move exactly one task from busiest to this_rq, as
4939 * part of active balancing operations within "domain".
4940 * Returns 1 if successful and 0 otherwise.
4942 * Called with both runqueues locked.
4944 static int move_one_task(struct lb_env *env)
4946 struct task_struct *p, *n;
4948 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4949 if (!can_migrate_task(p, env))
4954 * Right now, this is only the second place move_task()
4955 * is called, so we can safely collect move_task()
4956 * stats here rather than inside move_task().
4958 schedstat_inc(env->sd, lb_gained[env->idle]);
4964 static const unsigned int sched_nr_migrate_break = 32;
4967 * move_tasks tries to move up to imbalance weighted load from busiest to
4968 * this_rq, as part of a balancing operation within domain "sd".
4969 * Returns 1 if successful and 0 otherwise.
4971 * Called with both runqueues locked.
4973 static int move_tasks(struct lb_env *env)
4975 struct list_head *tasks = &env->src_rq->cfs_tasks;
4976 struct task_struct *p;
4980 if (env->imbalance <= 0)
4983 while (!list_empty(tasks)) {
4984 p = list_first_entry(tasks, struct task_struct, se.group_node);
4987 /* We've more or less seen every task there is, call it quits */
4988 if (env->loop > env->loop_max)
4991 /* take a breather every nr_migrate tasks */
4992 if (env->loop > env->loop_break) {
4993 env->loop_break += sched_nr_migrate_break;
4994 env->flags |= LBF_NEED_BREAK;
4998 if (!can_migrate_task(p, env))
5001 load = task_h_load(p);
5003 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5006 if ((load / 2) > env->imbalance)
5011 env->imbalance -= load;
5013 #ifdef CONFIG_PREEMPT
5015 * NEWIDLE balancing is a source of latency, so preemptible
5016 * kernels will stop after the first task is pulled to minimize
5017 * the critical section.
5019 if (env->idle == CPU_NEWLY_IDLE)
5024 * We only want to steal up to the prescribed amount of
5027 if (env->imbalance <= 0)
5032 list_move_tail(&p->se.group_node, tasks);
5036 * Right now, this is one of only two places move_task() is called,
5037 * so we can safely collect move_task() stats here rather than
5038 * inside move_task().
5040 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5045 #ifdef CONFIG_FAIR_GROUP_SCHED
5047 * update tg->load_weight by folding this cpu's load_avg
5049 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5051 struct sched_entity *se = tg->se[cpu];
5052 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5054 /* throttled entities do not contribute to load */
5055 if (throttled_hierarchy(cfs_rq))
5058 update_cfs_rq_blocked_load(cfs_rq, 1);
5061 update_entity_load_avg(se, 1);
5063 * We pivot on our runnable average having decayed to zero for
5064 * list removal. This generally implies that all our children
5065 * have also been removed (modulo rounding error or bandwidth
5066 * control); however, such cases are rare and we can fix these
5069 * TODO: fix up out-of-order children on enqueue.
5071 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5072 list_del_leaf_cfs_rq(cfs_rq);
5074 struct rq *rq = rq_of(cfs_rq);
5075 update_rq_runnable_avg(rq, rq->nr_running);
5079 static void update_blocked_averages(int cpu)
5081 struct rq *rq = cpu_rq(cpu);
5082 struct cfs_rq *cfs_rq;
5083 unsigned long flags;
5085 raw_spin_lock_irqsave(&rq->lock, flags);
5086 update_rq_clock(rq);
5088 * Iterates the task_group tree in a bottom up fashion, see
5089 * list_add_leaf_cfs_rq() for details.
5091 for_each_leaf_cfs_rq(rq, cfs_rq) {
5093 * Note: We may want to consider periodically releasing
5094 * rq->lock about these updates so that creating many task
5095 * groups does not result in continually extending hold time.
5097 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5100 raw_spin_unlock_irqrestore(&rq->lock, flags);
5104 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5105 * This needs to be done in a top-down fashion because the load of a child
5106 * group is a fraction of its parents load.
5108 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5110 struct rq *rq = rq_of(cfs_rq);
5111 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5112 unsigned long now = jiffies;
5115 if (cfs_rq->last_h_load_update == now)
5118 cfs_rq->h_load_next = NULL;
5119 for_each_sched_entity(se) {
5120 cfs_rq = cfs_rq_of(se);
5121 cfs_rq->h_load_next = se;
5122 if (cfs_rq->last_h_load_update == now)
5127 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5128 cfs_rq->last_h_load_update = now;
5131 while ((se = cfs_rq->h_load_next) != NULL) {
5132 load = cfs_rq->h_load;
5133 load = div64_ul(load * se->avg.load_avg_contrib,
5134 cfs_rq->runnable_load_avg + 1);
5135 cfs_rq = group_cfs_rq(se);
5136 cfs_rq->h_load = load;
5137 cfs_rq->last_h_load_update = now;
5141 static unsigned long task_h_load(struct task_struct *p)
5143 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5145 update_cfs_rq_h_load(cfs_rq);
5146 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5147 cfs_rq->runnable_load_avg + 1);
5150 static inline void update_blocked_averages(int cpu)
5154 static unsigned long task_h_load(struct task_struct *p)
5156 return p->se.avg.load_avg_contrib;
5160 /********** Helpers for find_busiest_group ************************/
5162 * sg_lb_stats - stats of a sched_group required for load_balancing
5164 struct sg_lb_stats {
5165 unsigned long avg_load; /*Avg load across the CPUs of the group */
5166 unsigned long group_load; /* Total load over the CPUs of the group */
5167 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5168 unsigned long load_per_task;
5169 unsigned long group_power;
5170 unsigned int sum_nr_running; /* Nr tasks running in the group */
5171 unsigned int group_capacity;
5172 unsigned int idle_cpus;
5173 unsigned int group_weight;
5174 int group_imb; /* Is there an imbalance in the group ? */
5175 int group_has_capacity; /* Is there extra capacity in the group? */
5176 #ifdef CONFIG_NUMA_BALANCING
5177 unsigned int nr_numa_running;
5178 unsigned int nr_preferred_running;
5183 * sd_lb_stats - Structure to store the statistics of a sched_domain
5184 * during load balancing.
5186 struct sd_lb_stats {
5187 struct sched_group *busiest; /* Busiest group in this sd */
5188 struct sched_group *local; /* Local group in this sd */
5189 unsigned long total_load; /* Total load of all groups in sd */
5190 unsigned long total_pwr; /* Total power of all groups in sd */
5191 unsigned long avg_load; /* Average load across all groups in sd */
5193 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5194 struct sg_lb_stats local_stat; /* Statistics of the local group */
5197 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5200 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5201 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5202 * We must however clear busiest_stat::avg_load because
5203 * update_sd_pick_busiest() reads this before assignment.
5205 *sds = (struct sd_lb_stats){
5217 * get_sd_load_idx - Obtain the load index for a given sched domain.
5218 * @sd: The sched_domain whose load_idx is to be obtained.
5219 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5221 * Return: The load index.
5223 static inline int get_sd_load_idx(struct sched_domain *sd,
5224 enum cpu_idle_type idle)
5230 load_idx = sd->busy_idx;
5233 case CPU_NEWLY_IDLE:
5234 load_idx = sd->newidle_idx;
5237 load_idx = sd->idle_idx;
5244 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5246 return SCHED_POWER_SCALE;
5249 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5251 return default_scale_freq_power(sd, cpu);
5254 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5256 unsigned long weight = sd->span_weight;
5257 unsigned long smt_gain = sd->smt_gain;
5264 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5266 return default_scale_smt_power(sd, cpu);
5269 static unsigned long scale_rt_power(int cpu)
5271 struct rq *rq = cpu_rq(cpu);
5272 u64 total, available, age_stamp, avg;
5275 * Since we're reading these variables without serialization make sure
5276 * we read them once before doing sanity checks on them.
5278 age_stamp = ACCESS_ONCE(rq->age_stamp);
5279 avg = ACCESS_ONCE(rq->rt_avg);
5281 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5283 if (unlikely(total < avg)) {
5284 /* Ensures that power won't end up being negative */
5287 available = total - avg;
5290 if (unlikely((s64)total < SCHED_POWER_SCALE))
5291 total = SCHED_POWER_SCALE;
5293 total >>= SCHED_POWER_SHIFT;
5295 return div_u64(available, total);
5298 static void update_cpu_power(struct sched_domain *sd, int cpu)
5300 unsigned long weight = sd->span_weight;
5301 unsigned long power = SCHED_POWER_SCALE;
5302 struct sched_group *sdg = sd->groups;
5304 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5305 if (sched_feat(ARCH_POWER))
5306 power *= arch_scale_smt_power(sd, cpu);
5308 power *= default_scale_smt_power(sd, cpu);
5310 power >>= SCHED_POWER_SHIFT;
5313 sdg->sgp->power_orig = power;
5315 if (sched_feat(ARCH_POWER))
5316 power *= arch_scale_freq_power(sd, cpu);
5318 power *= default_scale_freq_power(sd, cpu);
5320 power >>= SCHED_POWER_SHIFT;
5322 power *= scale_rt_power(cpu);
5323 power >>= SCHED_POWER_SHIFT;
5328 cpu_rq(cpu)->cpu_power = power;
5329 sdg->sgp->power = power;
5332 void update_group_power(struct sched_domain *sd, int cpu)
5334 struct sched_domain *child = sd->child;
5335 struct sched_group *group, *sdg = sd->groups;
5336 unsigned long power, power_orig;
5337 unsigned long interval;
5339 interval = msecs_to_jiffies(sd->balance_interval);
5340 interval = clamp(interval, 1UL, max_load_balance_interval);
5341 sdg->sgp->next_update = jiffies + interval;
5344 update_cpu_power(sd, cpu);
5348 power_orig = power = 0;
5350 if (child->flags & SD_OVERLAP) {
5352 * SD_OVERLAP domains cannot assume that child groups
5353 * span the current group.
5356 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5357 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5359 power_orig += sg->sgp->power_orig;
5360 power += sg->sgp->power;
5364 * !SD_OVERLAP domains can assume that child groups
5365 * span the current group.
5368 group = child->groups;
5370 power_orig += group->sgp->power_orig;
5371 power += group->sgp->power;
5372 group = group->next;
5373 } while (group != child->groups);
5376 sdg->sgp->power_orig = power_orig;
5377 sdg->sgp->power = power;
5381 * Try and fix up capacity for tiny siblings, this is needed when
5382 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5383 * which on its own isn't powerful enough.
5385 * See update_sd_pick_busiest() and check_asym_packing().
5388 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5391 * Only siblings can have significantly less than SCHED_POWER_SCALE
5393 if (!(sd->flags & SD_SHARE_CPUPOWER))
5397 * If ~90% of the cpu_power is still there, we're good.
5399 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5406 * Group imbalance indicates (and tries to solve) the problem where balancing
5407 * groups is inadequate due to tsk_cpus_allowed() constraints.
5409 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5410 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5413 * { 0 1 2 3 } { 4 5 6 7 }
5416 * If we were to balance group-wise we'd place two tasks in the first group and
5417 * two tasks in the second group. Clearly this is undesired as it will overload
5418 * cpu 3 and leave one of the cpus in the second group unused.
5420 * The current solution to this issue is detecting the skew in the first group
5421 * by noticing the lower domain failed to reach balance and had difficulty
5422 * moving tasks due to affinity constraints.
5424 * When this is so detected; this group becomes a candidate for busiest; see
5425 * update_sd_pick_busiest(). And calculate_imbalance() and
5426 * find_busiest_group() avoid some of the usual balance conditions to allow it
5427 * to create an effective group imbalance.
5429 * This is a somewhat tricky proposition since the next run might not find the
5430 * group imbalance and decide the groups need to be balanced again. A most
5431 * subtle and fragile situation.
5434 static inline int sg_imbalanced(struct sched_group *group)
5436 return group->sgp->imbalance;
5440 * Compute the group capacity.
5442 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5443 * first dividing out the smt factor and computing the actual number of cores
5444 * and limit power unit capacity with that.
5446 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5448 unsigned int capacity, smt, cpus;
5449 unsigned int power, power_orig;
5451 power = group->sgp->power;
5452 power_orig = group->sgp->power_orig;
5453 cpus = group->group_weight;
5455 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5456 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5457 capacity = cpus / smt; /* cores */
5459 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5461 capacity = fix_small_capacity(env->sd, group);
5467 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5468 * @env: The load balancing environment.
5469 * @group: sched_group whose statistics are to be updated.
5470 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5471 * @local_group: Does group contain this_cpu.
5472 * @sgs: variable to hold the statistics for this group.
5474 static inline void update_sg_lb_stats(struct lb_env *env,
5475 struct sched_group *group, int load_idx,
5476 int local_group, struct sg_lb_stats *sgs)
5478 unsigned long nr_running;
5482 memset(sgs, 0, sizeof(*sgs));
5484 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5485 struct rq *rq = cpu_rq(i);
5487 nr_running = rq->nr_running;
5489 /* Bias balancing toward cpus of our domain */
5491 load = target_load(i, load_idx);
5493 load = source_load(i, load_idx);
5495 sgs->group_load += load;
5496 sgs->sum_nr_running += nr_running;
5497 #ifdef CONFIG_NUMA_BALANCING
5498 sgs->nr_numa_running += rq->nr_numa_running;
5499 sgs->nr_preferred_running += rq->nr_preferred_running;
5501 sgs->sum_weighted_load += weighted_cpuload(i);
5506 /* Adjust by relative CPU power of the group */
5507 sgs->group_power = group->sgp->power;
5508 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5510 if (sgs->sum_nr_running)
5511 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5513 sgs->group_weight = group->group_weight;
5515 sgs->group_imb = sg_imbalanced(group);
5516 sgs->group_capacity = sg_capacity(env, group);
5518 if (sgs->group_capacity > sgs->sum_nr_running)
5519 sgs->group_has_capacity = 1;
5523 * update_sd_pick_busiest - return 1 on busiest group
5524 * @env: The load balancing environment.
5525 * @sds: sched_domain statistics
5526 * @sg: sched_group candidate to be checked for being the busiest
5527 * @sgs: sched_group statistics
5529 * Determine if @sg is a busier group than the previously selected
5532 * Return: %true if @sg is a busier group than the previously selected
5533 * busiest group. %false otherwise.
5535 static bool update_sd_pick_busiest(struct lb_env *env,
5536 struct sd_lb_stats *sds,
5537 struct sched_group *sg,
5538 struct sg_lb_stats *sgs)
5540 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5543 if (sgs->sum_nr_running > sgs->group_capacity)
5550 * ASYM_PACKING needs to move all the work to the lowest
5551 * numbered CPUs in the group, therefore mark all groups
5552 * higher than ourself as busy.
5554 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5555 env->dst_cpu < group_first_cpu(sg)) {
5559 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5566 #ifdef CONFIG_NUMA_BALANCING
5567 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5569 if (sgs->sum_nr_running > sgs->nr_numa_running)
5571 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5576 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5578 if (rq->nr_running > rq->nr_numa_running)
5580 if (rq->nr_running > rq->nr_preferred_running)
5585 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5590 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5594 #endif /* CONFIG_NUMA_BALANCING */
5597 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5598 * @env: The load balancing environment.
5599 * @sds: variable to hold the statistics for this sched_domain.
5601 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5603 struct sched_domain *child = env->sd->child;
5604 struct sched_group *sg = env->sd->groups;
5605 struct sg_lb_stats tmp_sgs;
5606 int load_idx, prefer_sibling = 0;
5608 if (child && child->flags & SD_PREFER_SIBLING)
5611 load_idx = get_sd_load_idx(env->sd, env->idle);
5614 struct sg_lb_stats *sgs = &tmp_sgs;
5617 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5620 sgs = &sds->local_stat;
5622 if (env->idle != CPU_NEWLY_IDLE ||
5623 time_after_eq(jiffies, sg->sgp->next_update))
5624 update_group_power(env->sd, env->dst_cpu);
5627 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5633 * In case the child domain prefers tasks go to siblings
5634 * first, lower the sg capacity to one so that we'll try
5635 * and move all the excess tasks away. We lower the capacity
5636 * of a group only if the local group has the capacity to fit
5637 * these excess tasks, i.e. nr_running < group_capacity. The
5638 * extra check prevents the case where you always pull from the
5639 * heaviest group when it is already under-utilized (possible
5640 * with a large weight task outweighs the tasks on the system).
5642 if (prefer_sibling && sds->local &&
5643 sds->local_stat.group_has_capacity)
5644 sgs->group_capacity = min(sgs->group_capacity, 1U);
5646 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5648 sds->busiest_stat = *sgs;
5652 /* Now, start updating sd_lb_stats */
5653 sds->total_load += sgs->group_load;
5654 sds->total_pwr += sgs->group_power;
5657 } while (sg != env->sd->groups);
5659 if (env->sd->flags & SD_NUMA)
5660 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5664 * check_asym_packing - Check to see if the group is packed into the
5667 * This is primarily intended to used at the sibling level. Some
5668 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5669 * case of POWER7, it can move to lower SMT modes only when higher
5670 * threads are idle. When in lower SMT modes, the threads will
5671 * perform better since they share less core resources. Hence when we
5672 * have idle threads, we want them to be the higher ones.
5674 * This packing function is run on idle threads. It checks to see if
5675 * the busiest CPU in this domain (core in the P7 case) has a higher
5676 * CPU number than the packing function is being run on. Here we are
5677 * assuming lower CPU number will be equivalent to lower a SMT thread
5680 * Return: 1 when packing is required and a task should be moved to
5681 * this CPU. The amount of the imbalance is returned in *imbalance.
5683 * @env: The load balancing environment.
5684 * @sds: Statistics of the sched_domain which is to be packed
5686 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5690 if (!(env->sd->flags & SD_ASYM_PACKING))
5696 busiest_cpu = group_first_cpu(sds->busiest);
5697 if (env->dst_cpu > busiest_cpu)
5700 env->imbalance = DIV_ROUND_CLOSEST(
5701 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5708 * fix_small_imbalance - Calculate the minor imbalance that exists
5709 * amongst the groups of a sched_domain, during
5711 * @env: The load balancing environment.
5712 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5715 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5717 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5718 unsigned int imbn = 2;
5719 unsigned long scaled_busy_load_per_task;
5720 struct sg_lb_stats *local, *busiest;
5722 local = &sds->local_stat;
5723 busiest = &sds->busiest_stat;
5725 if (!local->sum_nr_running)
5726 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5727 else if (busiest->load_per_task > local->load_per_task)
5730 scaled_busy_load_per_task =
5731 (busiest->load_per_task * SCHED_POWER_SCALE) /
5732 busiest->group_power;
5734 if (busiest->avg_load + scaled_busy_load_per_task >=
5735 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5736 env->imbalance = busiest->load_per_task;
5741 * OK, we don't have enough imbalance to justify moving tasks,
5742 * however we may be able to increase total CPU power used by
5746 pwr_now += busiest->group_power *
5747 min(busiest->load_per_task, busiest->avg_load);
5748 pwr_now += local->group_power *
5749 min(local->load_per_task, local->avg_load);
5750 pwr_now /= SCHED_POWER_SCALE;
5752 /* Amount of load we'd subtract */
5753 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5754 busiest->group_power;
5755 if (busiest->avg_load > tmp) {
5756 pwr_move += busiest->group_power *
5757 min(busiest->load_per_task,
5758 busiest->avg_load - tmp);
5761 /* Amount of load we'd add */
5762 if (busiest->avg_load * busiest->group_power <
5763 busiest->load_per_task * SCHED_POWER_SCALE) {
5764 tmp = (busiest->avg_load * busiest->group_power) /
5767 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5770 pwr_move += local->group_power *
5771 min(local->load_per_task, local->avg_load + tmp);
5772 pwr_move /= SCHED_POWER_SCALE;
5774 /* Move if we gain throughput */
5775 if (pwr_move > pwr_now)
5776 env->imbalance = busiest->load_per_task;
5780 * calculate_imbalance - Calculate the amount of imbalance present within the
5781 * groups of a given sched_domain during load balance.
5782 * @env: load balance environment
5783 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5785 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5787 unsigned long max_pull, load_above_capacity = ~0UL;
5788 struct sg_lb_stats *local, *busiest;
5790 local = &sds->local_stat;
5791 busiest = &sds->busiest_stat;
5793 if (busiest->group_imb) {
5795 * In the group_imb case we cannot rely on group-wide averages
5796 * to ensure cpu-load equilibrium, look at wider averages. XXX
5798 busiest->load_per_task =
5799 min(busiest->load_per_task, sds->avg_load);
5803 * In the presence of smp nice balancing, certain scenarios can have
5804 * max load less than avg load(as we skip the groups at or below
5805 * its cpu_power, while calculating max_load..)
5807 if (busiest->avg_load <= sds->avg_load ||
5808 local->avg_load >= sds->avg_load) {
5810 return fix_small_imbalance(env, sds);
5813 if (!busiest->group_imb) {
5815 * Don't want to pull so many tasks that a group would go idle.
5816 * Except of course for the group_imb case, since then we might
5817 * have to drop below capacity to reach cpu-load equilibrium.
5819 load_above_capacity =
5820 (busiest->sum_nr_running - busiest->group_capacity);
5822 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5823 load_above_capacity /= busiest->group_power;
5827 * We're trying to get all the cpus to the average_load, so we don't
5828 * want to push ourselves above the average load, nor do we wish to
5829 * reduce the max loaded cpu below the average load. At the same time,
5830 * we also don't want to reduce the group load below the group capacity
5831 * (so that we can implement power-savings policies etc). Thus we look
5832 * for the minimum possible imbalance.
5834 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5836 /* How much load to actually move to equalise the imbalance */
5837 env->imbalance = min(
5838 max_pull * busiest->group_power,
5839 (sds->avg_load - local->avg_load) * local->group_power
5840 ) / SCHED_POWER_SCALE;
5843 * if *imbalance is less than the average load per runnable task
5844 * there is no guarantee that any tasks will be moved so we'll have
5845 * a think about bumping its value to force at least one task to be
5848 if (env->imbalance < busiest->load_per_task)
5849 return fix_small_imbalance(env, sds);
5852 /******* find_busiest_group() helpers end here *********************/
5855 * find_busiest_group - Returns the busiest group within the sched_domain
5856 * if there is an imbalance. If there isn't an imbalance, and
5857 * the user has opted for power-savings, it returns a group whose
5858 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5859 * such a group exists.
5861 * Also calculates the amount of weighted load which should be moved
5862 * to restore balance.
5864 * @env: The load balancing environment.
5866 * Return: - The busiest group if imbalance exists.
5867 * - If no imbalance and user has opted for power-savings balance,
5868 * return the least loaded group whose CPUs can be
5869 * put to idle by rebalancing its tasks onto our group.
5871 static struct sched_group *find_busiest_group(struct lb_env *env)
5873 struct sg_lb_stats *local, *busiest;
5874 struct sd_lb_stats sds;
5876 init_sd_lb_stats(&sds);
5879 * Compute the various statistics relavent for load balancing at
5882 update_sd_lb_stats(env, &sds);
5883 local = &sds.local_stat;
5884 busiest = &sds.busiest_stat;
5886 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5887 check_asym_packing(env, &sds))
5890 /* There is no busy sibling group to pull tasks from */
5891 if (!sds.busiest || busiest->sum_nr_running == 0)
5894 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5897 * If the busiest group is imbalanced the below checks don't
5898 * work because they assume all things are equal, which typically
5899 * isn't true due to cpus_allowed constraints and the like.
5901 if (busiest->group_imb)
5904 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5905 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5906 !busiest->group_has_capacity)
5910 * If the local group is more busy than the selected busiest group
5911 * don't try and pull any tasks.
5913 if (local->avg_load >= busiest->avg_load)
5917 * Don't pull any tasks if this group is already above the domain
5920 if (local->avg_load >= sds.avg_load)
5923 if (env->idle == CPU_IDLE) {
5925 * This cpu is idle. If the busiest group load doesn't
5926 * have more tasks than the number of available cpu's and
5927 * there is no imbalance between this and busiest group
5928 * wrt to idle cpu's, it is balanced.
5930 if ((local->idle_cpus < busiest->idle_cpus) &&
5931 busiest->sum_nr_running <= busiest->group_weight)
5935 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5936 * imbalance_pct to be conservative.
5938 if (100 * busiest->avg_load <=
5939 env->sd->imbalance_pct * local->avg_load)
5944 /* Looks like there is an imbalance. Compute it */
5945 calculate_imbalance(env, &sds);
5954 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5956 static struct rq *find_busiest_queue(struct lb_env *env,
5957 struct sched_group *group)
5959 struct rq *busiest = NULL, *rq;
5960 unsigned long busiest_load = 0, busiest_power = 1;
5963 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5964 unsigned long power, capacity, wl;
5968 rt = fbq_classify_rq(rq);
5971 * We classify groups/runqueues into three groups:
5972 * - regular: there are !numa tasks
5973 * - remote: there are numa tasks that run on the 'wrong' node
5974 * - all: there is no distinction
5976 * In order to avoid migrating ideally placed numa tasks,
5977 * ignore those when there's better options.
5979 * If we ignore the actual busiest queue to migrate another
5980 * task, the next balance pass can still reduce the busiest
5981 * queue by moving tasks around inside the node.
5983 * If we cannot move enough load due to this classification
5984 * the next pass will adjust the group classification and
5985 * allow migration of more tasks.
5987 * Both cases only affect the total convergence complexity.
5989 if (rt > env->fbq_type)
5992 power = power_of(i);
5993 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5995 capacity = fix_small_capacity(env->sd, group);
5997 wl = weighted_cpuload(i);
6000 * When comparing with imbalance, use weighted_cpuload()
6001 * which is not scaled with the cpu power.
6003 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6007 * For the load comparisons with the other cpu's, consider
6008 * the weighted_cpuload() scaled with the cpu power, so that
6009 * the load can be moved away from the cpu that is potentially
6010 * running at a lower capacity.
6012 * Thus we're looking for max(wl_i / power_i), crosswise
6013 * multiplication to rid ourselves of the division works out
6014 * to: wl_i * power_j > wl_j * power_i; where j is our
6017 if (wl * busiest_power > busiest_load * power) {
6019 busiest_power = power;
6028 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6029 * so long as it is large enough.
6031 #define MAX_PINNED_INTERVAL 512
6033 /* Working cpumask for load_balance and load_balance_newidle. */
6034 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6036 static int need_active_balance(struct lb_env *env)
6038 struct sched_domain *sd = env->sd;
6040 if (env->idle == CPU_NEWLY_IDLE) {
6043 * ASYM_PACKING needs to force migrate tasks from busy but
6044 * higher numbered CPUs in order to pack all tasks in the
6045 * lowest numbered CPUs.
6047 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6051 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6054 static int active_load_balance_cpu_stop(void *data);
6056 static int should_we_balance(struct lb_env *env)
6058 struct sched_group *sg = env->sd->groups;
6059 struct cpumask *sg_cpus, *sg_mask;
6060 int cpu, balance_cpu = -1;
6063 * In the newly idle case, we will allow all the cpu's
6064 * to do the newly idle load balance.
6066 if (env->idle == CPU_NEWLY_IDLE)
6069 sg_cpus = sched_group_cpus(sg);
6070 sg_mask = sched_group_mask(sg);
6071 /* Try to find first idle cpu */
6072 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6073 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6080 if (balance_cpu == -1)
6081 balance_cpu = group_balance_cpu(sg);
6084 * First idle cpu or the first cpu(busiest) in this sched group
6085 * is eligible for doing load balancing at this and above domains.
6087 return balance_cpu == env->dst_cpu;
6091 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6092 * tasks if there is an imbalance.
6094 static int load_balance(int this_cpu, struct rq *this_rq,
6095 struct sched_domain *sd, enum cpu_idle_type idle,
6096 int *continue_balancing)
6098 int ld_moved, cur_ld_moved, active_balance = 0;
6099 struct sched_domain *sd_parent = sd->parent;
6100 struct sched_group *group;
6102 unsigned long flags;
6103 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6105 struct lb_env env = {
6107 .dst_cpu = this_cpu,
6109 .dst_grpmask = sched_group_cpus(sd->groups),
6111 .loop_break = sched_nr_migrate_break,
6117 * For NEWLY_IDLE load_balancing, we don't need to consider
6118 * other cpus in our group
6120 if (idle == CPU_NEWLY_IDLE)
6121 env.dst_grpmask = NULL;
6123 cpumask_copy(cpus, cpu_active_mask);
6125 schedstat_inc(sd, lb_count[idle]);
6128 if (!should_we_balance(&env)) {
6129 *continue_balancing = 0;
6133 group = find_busiest_group(&env);
6135 schedstat_inc(sd, lb_nobusyg[idle]);
6139 busiest = find_busiest_queue(&env, group);
6141 schedstat_inc(sd, lb_nobusyq[idle]);
6145 BUG_ON(busiest == env.dst_rq);
6147 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6150 if (busiest->nr_running > 1) {
6152 * Attempt to move tasks. If find_busiest_group has found
6153 * an imbalance but busiest->nr_running <= 1, the group is
6154 * still unbalanced. ld_moved simply stays zero, so it is
6155 * correctly treated as an imbalance.
6157 env.flags |= LBF_ALL_PINNED;
6158 env.src_cpu = busiest->cpu;
6159 env.src_rq = busiest;
6160 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6163 local_irq_save(flags);
6164 double_rq_lock(env.dst_rq, busiest);
6167 * cur_ld_moved - load moved in current iteration
6168 * ld_moved - cumulative load moved across iterations
6170 cur_ld_moved = move_tasks(&env);
6171 ld_moved += cur_ld_moved;
6172 double_rq_unlock(env.dst_rq, busiest);
6173 local_irq_restore(flags);
6176 * some other cpu did the load balance for us.
6178 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6179 resched_cpu(env.dst_cpu);
6181 if (env.flags & LBF_NEED_BREAK) {
6182 env.flags &= ~LBF_NEED_BREAK;
6187 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6188 * us and move them to an alternate dst_cpu in our sched_group
6189 * where they can run. The upper limit on how many times we
6190 * iterate on same src_cpu is dependent on number of cpus in our
6193 * This changes load balance semantics a bit on who can move
6194 * load to a given_cpu. In addition to the given_cpu itself
6195 * (or a ilb_cpu acting on its behalf where given_cpu is
6196 * nohz-idle), we now have balance_cpu in a position to move
6197 * load to given_cpu. In rare situations, this may cause
6198 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6199 * _independently_ and at _same_ time to move some load to
6200 * given_cpu) causing exceess load to be moved to given_cpu.
6201 * This however should not happen so much in practice and
6202 * moreover subsequent load balance cycles should correct the
6203 * excess load moved.
6205 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6207 /* Prevent to re-select dst_cpu via env's cpus */
6208 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6210 env.dst_rq = cpu_rq(env.new_dst_cpu);
6211 env.dst_cpu = env.new_dst_cpu;
6212 env.flags &= ~LBF_DST_PINNED;
6214 env.loop_break = sched_nr_migrate_break;
6217 * Go back to "more_balance" rather than "redo" since we
6218 * need to continue with same src_cpu.
6224 * We failed to reach balance because of affinity.
6227 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6229 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6230 *group_imbalance = 1;
6231 } else if (*group_imbalance)
6232 *group_imbalance = 0;
6235 /* All tasks on this runqueue were pinned by CPU affinity */
6236 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6237 cpumask_clear_cpu(cpu_of(busiest), cpus);
6238 if (!cpumask_empty(cpus)) {
6240 env.loop_break = sched_nr_migrate_break;
6248 schedstat_inc(sd, lb_failed[idle]);
6250 * Increment the failure counter only on periodic balance.
6251 * We do not want newidle balance, which can be very
6252 * frequent, pollute the failure counter causing
6253 * excessive cache_hot migrations and active balances.
6255 if (idle != CPU_NEWLY_IDLE)
6256 sd->nr_balance_failed++;
6258 if (need_active_balance(&env)) {
6259 raw_spin_lock_irqsave(&busiest->lock, flags);
6261 /* don't kick the active_load_balance_cpu_stop,
6262 * if the curr task on busiest cpu can't be
6265 if (!cpumask_test_cpu(this_cpu,
6266 tsk_cpus_allowed(busiest->curr))) {
6267 raw_spin_unlock_irqrestore(&busiest->lock,
6269 env.flags |= LBF_ALL_PINNED;
6270 goto out_one_pinned;
6274 * ->active_balance synchronizes accesses to
6275 * ->active_balance_work. Once set, it's cleared
6276 * only after active load balance is finished.
6278 if (!busiest->active_balance) {
6279 busiest->active_balance = 1;
6280 busiest->push_cpu = this_cpu;
6283 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6285 if (active_balance) {
6286 stop_one_cpu_nowait(cpu_of(busiest),
6287 active_load_balance_cpu_stop, busiest,
6288 &busiest->active_balance_work);
6292 * We've kicked active balancing, reset the failure
6295 sd->nr_balance_failed = sd->cache_nice_tries+1;
6298 sd->nr_balance_failed = 0;
6300 if (likely(!active_balance)) {
6301 /* We were unbalanced, so reset the balancing interval */
6302 sd->balance_interval = sd->min_interval;
6305 * If we've begun active balancing, start to back off. This
6306 * case may not be covered by the all_pinned logic if there
6307 * is only 1 task on the busy runqueue (because we don't call
6310 if (sd->balance_interval < sd->max_interval)
6311 sd->balance_interval *= 2;
6317 schedstat_inc(sd, lb_balanced[idle]);
6319 sd->nr_balance_failed = 0;
6322 /* tune up the balancing interval */
6323 if (((env.flags & LBF_ALL_PINNED) &&
6324 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6325 (sd->balance_interval < sd->max_interval))
6326 sd->balance_interval *= 2;
6334 * idle_balance is called by schedule() if this_cpu is about to become
6335 * idle. Attempts to pull tasks from other CPUs.
6337 void idle_balance(int this_cpu, struct rq *this_rq)
6339 struct sched_domain *sd;
6340 int pulled_task = 0;
6341 unsigned long next_balance = jiffies + HZ;
6344 this_rq->idle_stamp = rq_clock(this_rq);
6346 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6350 * Drop the rq->lock, but keep IRQ/preempt disabled.
6352 raw_spin_unlock(&this_rq->lock);
6354 update_blocked_averages(this_cpu);
6356 for_each_domain(this_cpu, sd) {
6357 unsigned long interval;
6358 int continue_balancing = 1;
6359 u64 t0, domain_cost;
6361 if (!(sd->flags & SD_LOAD_BALANCE))
6364 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6367 if (sd->flags & SD_BALANCE_NEWIDLE) {
6368 t0 = sched_clock_cpu(this_cpu);
6370 /* If we've pulled tasks over stop searching: */
6371 pulled_task = load_balance(this_cpu, this_rq,
6373 &continue_balancing);
6375 domain_cost = sched_clock_cpu(this_cpu) - t0;
6376 if (domain_cost > sd->max_newidle_lb_cost)
6377 sd->max_newidle_lb_cost = domain_cost;
6379 curr_cost += domain_cost;
6382 interval = msecs_to_jiffies(sd->balance_interval);
6383 if (time_after(next_balance, sd->last_balance + interval))
6384 next_balance = sd->last_balance + interval;
6386 this_rq->idle_stamp = 0;
6392 raw_spin_lock(&this_rq->lock);
6394 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6396 * We are going idle. next_balance may be set based on
6397 * a busy processor. So reset next_balance.
6399 this_rq->next_balance = next_balance;
6402 if (curr_cost > this_rq->max_idle_balance_cost)
6403 this_rq->max_idle_balance_cost = curr_cost;
6407 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6408 * running tasks off the busiest CPU onto idle CPUs. It requires at
6409 * least 1 task to be running on each physical CPU where possible, and
6410 * avoids physical / logical imbalances.
6412 static int active_load_balance_cpu_stop(void *data)
6414 struct rq *busiest_rq = data;
6415 int busiest_cpu = cpu_of(busiest_rq);
6416 int target_cpu = busiest_rq->push_cpu;
6417 struct rq *target_rq = cpu_rq(target_cpu);
6418 struct sched_domain *sd;
6420 raw_spin_lock_irq(&busiest_rq->lock);
6422 /* make sure the requested cpu hasn't gone down in the meantime */
6423 if (unlikely(busiest_cpu != smp_processor_id() ||
6424 !busiest_rq->active_balance))
6427 /* Is there any task to move? */
6428 if (busiest_rq->nr_running <= 1)
6432 * This condition is "impossible", if it occurs
6433 * we need to fix it. Originally reported by
6434 * Bjorn Helgaas on a 128-cpu setup.
6436 BUG_ON(busiest_rq == target_rq);
6438 /* move a task from busiest_rq to target_rq */
6439 double_lock_balance(busiest_rq, target_rq);
6441 /* Search for an sd spanning us and the target CPU. */
6443 for_each_domain(target_cpu, sd) {
6444 if ((sd->flags & SD_LOAD_BALANCE) &&
6445 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6450 struct lb_env env = {
6452 .dst_cpu = target_cpu,
6453 .dst_rq = target_rq,
6454 .src_cpu = busiest_rq->cpu,
6455 .src_rq = busiest_rq,
6459 schedstat_inc(sd, alb_count);
6461 if (move_one_task(&env))
6462 schedstat_inc(sd, alb_pushed);
6464 schedstat_inc(sd, alb_failed);
6467 double_unlock_balance(busiest_rq, target_rq);
6469 busiest_rq->active_balance = 0;
6470 raw_spin_unlock_irq(&busiest_rq->lock);
6474 #ifdef CONFIG_NO_HZ_COMMON
6476 * idle load balancing details
6477 * - When one of the busy CPUs notice that there may be an idle rebalancing
6478 * needed, they will kick the idle load balancer, which then does idle
6479 * load balancing for all the idle CPUs.
6482 cpumask_var_t idle_cpus_mask;
6484 unsigned long next_balance; /* in jiffy units */
6485 } nohz ____cacheline_aligned;
6487 static inline int find_new_ilb(int call_cpu)
6489 int ilb = cpumask_first(nohz.idle_cpus_mask);
6491 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6498 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6499 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6500 * CPU (if there is one).
6502 static void nohz_balancer_kick(int cpu)
6506 nohz.next_balance++;
6508 ilb_cpu = find_new_ilb(cpu);
6510 if (ilb_cpu >= nr_cpu_ids)
6513 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6516 * Use smp_send_reschedule() instead of resched_cpu().
6517 * This way we generate a sched IPI on the target cpu which
6518 * is idle. And the softirq performing nohz idle load balance
6519 * will be run before returning from the IPI.
6521 smp_send_reschedule(ilb_cpu);
6525 static inline void nohz_balance_exit_idle(int cpu)
6527 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6528 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6529 atomic_dec(&nohz.nr_cpus);
6530 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6534 static inline void set_cpu_sd_state_busy(void)
6536 struct sched_domain *sd;
6537 int cpu = smp_processor_id();
6540 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6542 if (!sd || !sd->nohz_idle)
6546 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6551 void set_cpu_sd_state_idle(void)
6553 struct sched_domain *sd;
6554 int cpu = smp_processor_id();
6557 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6559 if (!sd || sd->nohz_idle)
6563 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6569 * This routine will record that the cpu is going idle with tick stopped.
6570 * This info will be used in performing idle load balancing in the future.
6572 void nohz_balance_enter_idle(int cpu)
6575 * If this cpu is going down, then nothing needs to be done.
6577 if (!cpu_active(cpu))
6580 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6583 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6584 atomic_inc(&nohz.nr_cpus);
6585 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6588 static int sched_ilb_notifier(struct notifier_block *nfb,
6589 unsigned long action, void *hcpu)
6591 switch (action & ~CPU_TASKS_FROZEN) {
6593 nohz_balance_exit_idle(smp_processor_id());
6601 static DEFINE_SPINLOCK(balancing);
6604 * Scale the max load_balance interval with the number of CPUs in the system.
6605 * This trades load-balance latency on larger machines for less cross talk.
6607 void update_max_interval(void)
6609 max_load_balance_interval = HZ*num_online_cpus()/10;
6613 * It checks each scheduling domain to see if it is due to be balanced,
6614 * and initiates a balancing operation if so.
6616 * Balancing parameters are set up in init_sched_domains.
6618 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6620 int continue_balancing = 1;
6621 struct rq *rq = cpu_rq(cpu);
6622 unsigned long interval;
6623 struct sched_domain *sd;
6624 /* Earliest time when we have to do rebalance again */
6625 unsigned long next_balance = jiffies + 60*HZ;
6626 int update_next_balance = 0;
6627 int need_serialize, need_decay = 0;
6630 update_blocked_averages(cpu);
6633 for_each_domain(cpu, sd) {
6635 * Decay the newidle max times here because this is a regular
6636 * visit to all the domains. Decay ~1% per second.
6638 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6639 sd->max_newidle_lb_cost =
6640 (sd->max_newidle_lb_cost * 253) / 256;
6641 sd->next_decay_max_lb_cost = jiffies + HZ;
6644 max_cost += sd->max_newidle_lb_cost;
6646 if (!(sd->flags & SD_LOAD_BALANCE))
6650 * Stop the load balance at this level. There is another
6651 * CPU in our sched group which is doing load balancing more
6654 if (!continue_balancing) {
6660 interval = sd->balance_interval;
6661 if (idle != CPU_IDLE)
6662 interval *= sd->busy_factor;
6664 /* scale ms to jiffies */
6665 interval = msecs_to_jiffies(interval);
6666 interval = clamp(interval, 1UL, max_load_balance_interval);
6668 need_serialize = sd->flags & SD_SERIALIZE;
6670 if (need_serialize) {
6671 if (!spin_trylock(&balancing))
6675 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6676 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6678 * The LBF_DST_PINNED logic could have changed
6679 * env->dst_cpu, so we can't know our idle
6680 * state even if we migrated tasks. Update it.
6682 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6684 sd->last_balance = jiffies;
6687 spin_unlock(&balancing);
6689 if (time_after(next_balance, sd->last_balance + interval)) {
6690 next_balance = sd->last_balance + interval;
6691 update_next_balance = 1;
6696 * Ensure the rq-wide value also decays but keep it at a
6697 * reasonable floor to avoid funnies with rq->avg_idle.
6699 rq->max_idle_balance_cost =
6700 max((u64)sysctl_sched_migration_cost, max_cost);
6705 * next_balance will be updated only when there is a need.
6706 * When the cpu is attached to null domain for ex, it will not be
6709 if (likely(update_next_balance))
6710 rq->next_balance = next_balance;
6713 #ifdef CONFIG_NO_HZ_COMMON
6715 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6716 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6718 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6720 struct rq *this_rq = cpu_rq(this_cpu);
6724 if (idle != CPU_IDLE ||
6725 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6728 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6729 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6733 * If this cpu gets work to do, stop the load balancing
6734 * work being done for other cpus. Next load
6735 * balancing owner will pick it up.
6740 rq = cpu_rq(balance_cpu);
6742 raw_spin_lock_irq(&rq->lock);
6743 update_rq_clock(rq);
6744 update_idle_cpu_load(rq);
6745 raw_spin_unlock_irq(&rq->lock);
6747 rebalance_domains(balance_cpu, CPU_IDLE);
6749 if (time_after(this_rq->next_balance, rq->next_balance))
6750 this_rq->next_balance = rq->next_balance;
6752 nohz.next_balance = this_rq->next_balance;
6754 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6758 * Current heuristic for kicking the idle load balancer in the presence
6759 * of an idle cpu is the system.
6760 * - This rq has more than one task.
6761 * - At any scheduler domain level, this cpu's scheduler group has multiple
6762 * busy cpu's exceeding the group's power.
6763 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6764 * domain span are idle.
6766 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6768 unsigned long now = jiffies;
6769 struct sched_domain *sd;
6770 struct sched_group_power *sgp;
6773 if (unlikely(idle_cpu(cpu)))
6777 * We may be recently in ticked or tickless idle mode. At the first
6778 * busy tick after returning from idle, we will update the busy stats.
6780 set_cpu_sd_state_busy();
6781 nohz_balance_exit_idle(cpu);
6784 * None are in tickless mode and hence no need for NOHZ idle load
6787 if (likely(!atomic_read(&nohz.nr_cpus)))
6790 if (time_before(now, nohz.next_balance))
6793 if (rq->nr_running >= 2)
6797 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6800 sgp = sd->groups->sgp;
6801 nr_busy = atomic_read(&sgp->nr_busy_cpus);
6804 goto need_kick_unlock;
6807 sd = rcu_dereference(per_cpu(sd_asym, cpu));
6809 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
6810 sched_domain_span(sd)) < cpu))
6811 goto need_kick_unlock;
6822 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6826 * run_rebalance_domains is triggered when needed from the scheduler tick.
6827 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6829 static void run_rebalance_domains(struct softirq_action *h)
6831 int this_cpu = smp_processor_id();
6832 struct rq *this_rq = cpu_rq(this_cpu);
6833 enum cpu_idle_type idle = this_rq->idle_balance ?
6834 CPU_IDLE : CPU_NOT_IDLE;
6836 rebalance_domains(this_cpu, idle);
6839 * If this cpu has a pending nohz_balance_kick, then do the
6840 * balancing on behalf of the other idle cpus whose ticks are
6843 nohz_idle_balance(this_cpu, idle);
6846 static inline int on_null_domain(int cpu)
6848 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6852 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6854 void trigger_load_balance(struct rq *rq, int cpu)
6856 /* Don't need to rebalance while attached to NULL domain */
6857 if (time_after_eq(jiffies, rq->next_balance) &&
6858 likely(!on_null_domain(cpu)))
6859 raise_softirq(SCHED_SOFTIRQ);
6860 #ifdef CONFIG_NO_HZ_COMMON
6861 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6862 nohz_balancer_kick(cpu);
6866 static void rq_online_fair(struct rq *rq)
6871 static void rq_offline_fair(struct rq *rq)
6875 /* Ensure any throttled groups are reachable by pick_next_task */
6876 unthrottle_offline_cfs_rqs(rq);
6879 #endif /* CONFIG_SMP */
6882 * scheduler tick hitting a task of our scheduling class:
6884 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6886 struct cfs_rq *cfs_rq;
6887 struct sched_entity *se = &curr->se;
6889 for_each_sched_entity(se) {
6890 cfs_rq = cfs_rq_of(se);
6891 entity_tick(cfs_rq, se, queued);
6894 if (numabalancing_enabled)
6895 task_tick_numa(rq, curr);
6897 update_rq_runnable_avg(rq, 1);
6901 * called on fork with the child task as argument from the parent's context
6902 * - child not yet on the tasklist
6903 * - preemption disabled
6905 static void task_fork_fair(struct task_struct *p)
6907 struct cfs_rq *cfs_rq;
6908 struct sched_entity *se = &p->se, *curr;
6909 int this_cpu = smp_processor_id();
6910 struct rq *rq = this_rq();
6911 unsigned long flags;
6913 raw_spin_lock_irqsave(&rq->lock, flags);
6915 update_rq_clock(rq);
6917 cfs_rq = task_cfs_rq(current);
6918 curr = cfs_rq->curr;
6921 * Not only the cpu but also the task_group of the parent might have
6922 * been changed after parent->se.parent,cfs_rq were copied to
6923 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6924 * of child point to valid ones.
6927 __set_task_cpu(p, this_cpu);
6930 update_curr(cfs_rq);
6933 se->vruntime = curr->vruntime;
6934 place_entity(cfs_rq, se, 1);
6936 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6938 * Upon rescheduling, sched_class::put_prev_task() will place
6939 * 'current' within the tree based on its new key value.
6941 swap(curr->vruntime, se->vruntime);
6942 resched_task(rq->curr);
6945 se->vruntime -= cfs_rq->min_vruntime;
6947 raw_spin_unlock_irqrestore(&rq->lock, flags);
6951 * Priority of the task has changed. Check to see if we preempt
6955 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6961 * Reschedule if we are currently running on this runqueue and
6962 * our priority decreased, or if we are not currently running on
6963 * this runqueue and our priority is higher than the current's
6965 if (rq->curr == p) {
6966 if (p->prio > oldprio)
6967 resched_task(rq->curr);
6969 check_preempt_curr(rq, p, 0);
6972 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6974 struct sched_entity *se = &p->se;
6975 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6978 * Ensure the task's vruntime is normalized, so that when its
6979 * switched back to the fair class the enqueue_entity(.flags=0) will
6980 * do the right thing.
6982 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6983 * have normalized the vruntime, if it was !on_rq, then only when
6984 * the task is sleeping will it still have non-normalized vruntime.
6986 if (!se->on_rq && p->state != TASK_RUNNING) {
6988 * Fix up our vruntime so that the current sleep doesn't
6989 * cause 'unlimited' sleep bonus.
6991 place_entity(cfs_rq, se, 0);
6992 se->vruntime -= cfs_rq->min_vruntime;
6997 * Remove our load from contribution when we leave sched_fair
6998 * and ensure we don't carry in an old decay_count if we
7001 if (se->avg.decay_count) {
7002 __synchronize_entity_decay(se);
7003 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7009 * We switched to the sched_fair class.
7011 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7017 * We were most likely switched from sched_rt, so
7018 * kick off the schedule if running, otherwise just see
7019 * if we can still preempt the current task.
7022 resched_task(rq->curr);
7024 check_preempt_curr(rq, p, 0);
7027 /* Account for a task changing its policy or group.
7029 * This routine is mostly called to set cfs_rq->curr field when a task
7030 * migrates between groups/classes.
7032 static void set_curr_task_fair(struct rq *rq)
7034 struct sched_entity *se = &rq->curr->se;
7036 for_each_sched_entity(se) {
7037 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7039 set_next_entity(cfs_rq, se);
7040 /* ensure bandwidth has been allocated on our new cfs_rq */
7041 account_cfs_rq_runtime(cfs_rq, 0);
7045 void init_cfs_rq(struct cfs_rq *cfs_rq)
7047 cfs_rq->tasks_timeline = RB_ROOT;
7048 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7049 #ifndef CONFIG_64BIT
7050 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7053 atomic64_set(&cfs_rq->decay_counter, 1);
7054 atomic_long_set(&cfs_rq->removed_load, 0);
7058 #ifdef CONFIG_FAIR_GROUP_SCHED
7059 static void task_move_group_fair(struct task_struct *p, int on_rq)
7061 struct cfs_rq *cfs_rq;
7063 * If the task was not on the rq at the time of this cgroup movement
7064 * it must have been asleep, sleeping tasks keep their ->vruntime
7065 * absolute on their old rq until wakeup (needed for the fair sleeper
7066 * bonus in place_entity()).
7068 * If it was on the rq, we've just 'preempted' it, which does convert
7069 * ->vruntime to a relative base.
7071 * Make sure both cases convert their relative position when migrating
7072 * to another cgroup's rq. This does somewhat interfere with the
7073 * fair sleeper stuff for the first placement, but who cares.
7076 * When !on_rq, vruntime of the task has usually NOT been normalized.
7077 * But there are some cases where it has already been normalized:
7079 * - Moving a forked child which is waiting for being woken up by
7080 * wake_up_new_task().
7081 * - Moving a task which has been woken up by try_to_wake_up() and
7082 * waiting for actually being woken up by sched_ttwu_pending().
7084 * To prevent boost or penalty in the new cfs_rq caused by delta
7085 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7087 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7091 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7092 set_task_rq(p, task_cpu(p));
7094 cfs_rq = cfs_rq_of(&p->se);
7095 p->se.vruntime += cfs_rq->min_vruntime;
7098 * migrate_task_rq_fair() will have removed our previous
7099 * contribution, but we must synchronize for ongoing future
7102 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7103 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7108 void free_fair_sched_group(struct task_group *tg)
7112 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7114 for_each_possible_cpu(i) {
7116 kfree(tg->cfs_rq[i]);
7125 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7127 struct cfs_rq *cfs_rq;
7128 struct sched_entity *se;
7131 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7134 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7138 tg->shares = NICE_0_LOAD;
7140 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7142 for_each_possible_cpu(i) {
7143 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7144 GFP_KERNEL, cpu_to_node(i));
7148 se = kzalloc_node(sizeof(struct sched_entity),
7149 GFP_KERNEL, cpu_to_node(i));
7153 init_cfs_rq(cfs_rq);
7154 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7165 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7167 struct rq *rq = cpu_rq(cpu);
7168 unsigned long flags;
7171 * Only empty task groups can be destroyed; so we can speculatively
7172 * check on_list without danger of it being re-added.
7174 if (!tg->cfs_rq[cpu]->on_list)
7177 raw_spin_lock_irqsave(&rq->lock, flags);
7178 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7179 raw_spin_unlock_irqrestore(&rq->lock, flags);
7182 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7183 struct sched_entity *se, int cpu,
7184 struct sched_entity *parent)
7186 struct rq *rq = cpu_rq(cpu);
7190 init_cfs_rq_runtime(cfs_rq);
7192 tg->cfs_rq[cpu] = cfs_rq;
7195 /* se could be NULL for root_task_group */
7200 se->cfs_rq = &rq->cfs;
7202 se->cfs_rq = parent->my_q;
7205 /* guarantee group entities always have weight */
7206 update_load_set(&se->load, NICE_0_LOAD);
7207 se->parent = parent;
7210 static DEFINE_MUTEX(shares_mutex);
7212 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7215 unsigned long flags;
7218 * We can't change the weight of the root cgroup.
7223 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7225 mutex_lock(&shares_mutex);
7226 if (tg->shares == shares)
7229 tg->shares = shares;
7230 for_each_possible_cpu(i) {
7231 struct rq *rq = cpu_rq(i);
7232 struct sched_entity *se;
7235 /* Propagate contribution to hierarchy */
7236 raw_spin_lock_irqsave(&rq->lock, flags);
7238 /* Possible calls to update_curr() need rq clock */
7239 update_rq_clock(rq);
7240 for_each_sched_entity(se)
7241 update_cfs_shares(group_cfs_rq(se));
7242 raw_spin_unlock_irqrestore(&rq->lock, flags);
7246 mutex_unlock(&shares_mutex);
7249 #else /* CONFIG_FAIR_GROUP_SCHED */
7251 void free_fair_sched_group(struct task_group *tg) { }
7253 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7258 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7260 #endif /* CONFIG_FAIR_GROUP_SCHED */
7263 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7265 struct sched_entity *se = &task->se;
7266 unsigned int rr_interval = 0;
7269 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7272 if (rq->cfs.load.weight)
7273 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7279 * All the scheduling class methods:
7281 const struct sched_class fair_sched_class = {
7282 .next = &idle_sched_class,
7283 .enqueue_task = enqueue_task_fair,
7284 .dequeue_task = dequeue_task_fair,
7285 .yield_task = yield_task_fair,
7286 .yield_to_task = yield_to_task_fair,
7288 .check_preempt_curr = check_preempt_wakeup,
7290 .pick_next_task = pick_next_task_fair,
7291 .put_prev_task = put_prev_task_fair,
7294 .select_task_rq = select_task_rq_fair,
7295 .migrate_task_rq = migrate_task_rq_fair,
7297 .rq_online = rq_online_fair,
7298 .rq_offline = rq_offline_fair,
7300 .task_waking = task_waking_fair,
7303 .set_curr_task = set_curr_task_fair,
7304 .task_tick = task_tick_fair,
7305 .task_fork = task_fork_fair,
7307 .prio_changed = prio_changed_fair,
7308 .switched_from = switched_from_fair,
7309 .switched_to = switched_to_fair,
7311 .get_rr_interval = get_rr_interval_fair,
7313 #ifdef CONFIG_FAIR_GROUP_SCHED
7314 .task_move_group = task_move_group_fair,
7318 #ifdef CONFIG_SCHED_DEBUG
7319 void print_cfs_stats(struct seq_file *m, int cpu)
7321 struct cfs_rq *cfs_rq;
7324 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7325 print_cfs_rq(m, cpu, cfs_rq);
7330 __init void init_sched_fair_class(void)
7333 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7335 #ifdef CONFIG_NO_HZ_COMMON
7336 nohz.next_balance = jiffies;
7337 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7338 cpu_notifier(sched_ilb_notifier, 0);