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 atomic_long_t total_faults;
920 atomic_long_t 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 atomic_long_read(&p->numa_group->faults[2*nid]) +
948 atomic_long_read(&p->numa_group->faults[2*nid+1]);
952 * These return the fraction of accesses done by a particular task, or
953 * task group, on a particular numa node. The group weight is given a
954 * larger multiplier, in order to group tasks together that are almost
955 * evenly spread out between numa nodes.
957 static inline unsigned long task_weight(struct task_struct *p, int nid)
959 unsigned long total_faults;
964 total_faults = p->total_numa_faults;
969 return 1000 * task_faults(p, nid) / total_faults;
972 static inline unsigned long group_weight(struct task_struct *p, int nid)
974 unsigned long total_faults;
979 total_faults = atomic_long_read(&p->numa_group->total_faults);
984 return 1000 * group_faults(p, nid) / total_faults;
987 static unsigned long weighted_cpuload(const int cpu);
988 static unsigned long source_load(int cpu, int type);
989 static unsigned long target_load(int cpu, int type);
990 static unsigned long power_of(int cpu);
991 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
993 /* Cached statistics for all CPUs within a node */
995 unsigned long nr_running;
998 /* Total compute capacity of CPUs on a node */
1001 /* Approximate capacity in terms of runnable tasks on a node */
1002 unsigned long capacity;
1007 * XXX borrowed from update_sg_lb_stats
1009 static void update_numa_stats(struct numa_stats *ns, int nid)
1013 memset(ns, 0, sizeof(*ns));
1014 for_each_cpu(cpu, cpumask_of_node(nid)) {
1015 struct rq *rq = cpu_rq(cpu);
1017 ns->nr_running += rq->nr_running;
1018 ns->load += weighted_cpuload(cpu);
1019 ns->power += power_of(cpu);
1022 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1023 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1024 ns->has_capacity = (ns->nr_running < ns->capacity);
1027 struct task_numa_env {
1028 struct task_struct *p;
1030 int src_cpu, src_nid;
1031 int dst_cpu, dst_nid;
1033 struct numa_stats src_stats, dst_stats;
1035 int imbalance_pct, idx;
1037 struct task_struct *best_task;
1042 static void task_numa_assign(struct task_numa_env *env,
1043 struct task_struct *p, long imp)
1046 put_task_struct(env->best_task);
1051 env->best_imp = imp;
1052 env->best_cpu = env->dst_cpu;
1056 * This checks if the overall compute and NUMA accesses of the system would
1057 * be improved if the source tasks was migrated to the target dst_cpu taking
1058 * into account that it might be best if task running on the dst_cpu should
1059 * be exchanged with the source task
1061 static void task_numa_compare(struct task_numa_env *env,
1062 long taskimp, long groupimp)
1064 struct rq *src_rq = cpu_rq(env->src_cpu);
1065 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1066 struct task_struct *cur;
1067 long dst_load, src_load;
1069 long imp = (groupimp > 0) ? groupimp : taskimp;
1072 cur = ACCESS_ONCE(dst_rq->curr);
1073 if (cur->pid == 0) /* idle */
1077 * "imp" is the fault differential for the source task between the
1078 * source and destination node. Calculate the total differential for
1079 * the source task and potential destination task. The more negative
1080 * the value is, the more rmeote accesses that would be expected to
1081 * be incurred if the tasks were swapped.
1084 /* Skip this swap candidate if cannot move to the source cpu */
1085 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1089 * If dst and source tasks are in the same NUMA group, or not
1090 * in any group then look only at task weights.
1092 if (cur->numa_group == env->p->numa_group) {
1093 imp = taskimp + task_weight(cur, env->src_nid) -
1094 task_weight(cur, env->dst_nid);
1096 * Add some hysteresis to prevent swapping the
1097 * tasks within a group over tiny differences.
1099 if (cur->numa_group)
1103 * Compare the group weights. If a task is all by
1104 * itself (not part of a group), use the task weight
1107 if (env->p->numa_group)
1112 if (cur->numa_group)
1113 imp += group_weight(cur, env->src_nid) -
1114 group_weight(cur, env->dst_nid);
1116 imp += task_weight(cur, env->src_nid) -
1117 task_weight(cur, env->dst_nid);
1121 if (imp < env->best_imp)
1125 /* Is there capacity at our destination? */
1126 if (env->src_stats.has_capacity &&
1127 !env->dst_stats.has_capacity)
1133 /* Balance doesn't matter much if we're running a task per cpu */
1134 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1138 * In the overloaded case, try and keep the load balanced.
1141 dst_load = env->dst_stats.load;
1142 src_load = env->src_stats.load;
1144 /* XXX missing power terms */
1145 load = task_h_load(env->p);
1150 load = task_h_load(cur);
1155 /* make src_load the smaller */
1156 if (dst_load < src_load)
1157 swap(dst_load, src_load);
1159 if (src_load * env->imbalance_pct < dst_load * 100)
1163 task_numa_assign(env, cur, imp);
1168 static void task_numa_find_cpu(struct task_numa_env *env,
1169 long taskimp, long groupimp)
1173 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1174 /* Skip this CPU if the source task cannot migrate */
1175 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1179 task_numa_compare(env, taskimp, groupimp);
1183 static int task_numa_migrate(struct task_struct *p)
1185 struct task_numa_env env = {
1188 .src_cpu = task_cpu(p),
1189 .src_nid = task_node(p),
1191 .imbalance_pct = 112,
1197 struct sched_domain *sd;
1198 unsigned long taskweight, groupweight;
1200 long taskimp, groupimp;
1203 * Pick the lowest SD_NUMA domain, as that would have the smallest
1204 * imbalance and would be the first to start moving tasks about.
1206 * And we want to avoid any moving of tasks about, as that would create
1207 * random movement of tasks -- counter the numa conditions we're trying
1211 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1212 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1215 taskweight = task_weight(p, env.src_nid);
1216 groupweight = group_weight(p, env.src_nid);
1217 update_numa_stats(&env.src_stats, env.src_nid);
1218 env.dst_nid = p->numa_preferred_nid;
1219 taskimp = task_weight(p, env.dst_nid) - taskweight;
1220 groupimp = group_weight(p, env.dst_nid) - groupweight;
1221 update_numa_stats(&env.dst_stats, env.dst_nid);
1223 /* If the preferred nid has capacity, try to use it. */
1224 if (env.dst_stats.has_capacity)
1225 task_numa_find_cpu(&env, taskimp, groupimp);
1227 /* No space available on the preferred nid. Look elsewhere. */
1228 if (env.best_cpu == -1) {
1229 for_each_online_node(nid) {
1230 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1233 /* Only consider nodes where both task and groups benefit */
1234 taskimp = task_weight(p, nid) - taskweight;
1235 groupimp = group_weight(p, nid) - groupweight;
1236 if (taskimp < 0 && groupimp < 0)
1240 update_numa_stats(&env.dst_stats, env.dst_nid);
1241 task_numa_find_cpu(&env, taskimp, groupimp);
1245 /* No better CPU than the current one was found. */
1246 if (env.best_cpu == -1)
1249 sched_setnuma(p, env.dst_nid);
1252 * Reset the scan period if the task is being rescheduled on an
1253 * alternative node to recheck if the tasks is now properly placed.
1255 p->numa_scan_period = task_scan_min(p);
1257 if (env.best_task == NULL) {
1258 int ret = migrate_task_to(p, env.best_cpu);
1262 ret = migrate_swap(p, env.best_task);
1263 put_task_struct(env.best_task);
1267 /* Attempt to migrate a task to a CPU on the preferred node. */
1268 static void numa_migrate_preferred(struct task_struct *p)
1270 /* Success if task is already running on preferred CPU */
1271 p->numa_migrate_retry = 0;
1272 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid)
1275 /* This task has no NUMA fault statistics yet */
1276 if (unlikely(p->numa_preferred_nid == -1))
1279 /* Otherwise, try migrate to a CPU on the preferred node */
1280 if (task_numa_migrate(p) != 0)
1281 p->numa_migrate_retry = jiffies + HZ*5;
1285 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1286 * increments. The more local the fault statistics are, the higher the scan
1287 * period will be for the next scan window. If local/remote ratio is below
1288 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1289 * scan period will decrease
1291 #define NUMA_PERIOD_SLOTS 10
1292 #define NUMA_PERIOD_THRESHOLD 3
1295 * Increase the scan period (slow down scanning) if the majority of
1296 * our memory is already on our local node, or if the majority of
1297 * the page accesses are shared with other processes.
1298 * Otherwise, decrease the scan period.
1300 static void update_task_scan_period(struct task_struct *p,
1301 unsigned long shared, unsigned long private)
1303 unsigned int period_slot;
1307 unsigned long remote = p->numa_faults_locality[0];
1308 unsigned long local = p->numa_faults_locality[1];
1311 * If there were no record hinting faults then either the task is
1312 * completely idle or all activity is areas that are not of interest
1313 * to automatic numa balancing. Scan slower
1315 if (local + shared == 0) {
1316 p->numa_scan_period = min(p->numa_scan_period_max,
1317 p->numa_scan_period << 1);
1319 p->mm->numa_next_scan = jiffies +
1320 msecs_to_jiffies(p->numa_scan_period);
1326 * Prepare to scale scan period relative to the current period.
1327 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1328 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1329 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1331 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1332 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1333 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1334 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1337 diff = slot * period_slot;
1339 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1342 * Scale scan rate increases based on sharing. There is an
1343 * inverse relationship between the degree of sharing and
1344 * the adjustment made to the scanning period. Broadly
1345 * speaking the intent is that there is little point
1346 * scanning faster if shared accesses dominate as it may
1347 * simply bounce migrations uselessly
1349 period_slot = DIV_ROUND_UP(diff, NUMA_PERIOD_SLOTS);
1350 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1351 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1354 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1355 task_scan_min(p), task_scan_max(p));
1356 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1359 static void task_numa_placement(struct task_struct *p)
1361 int seq, nid, max_nid = -1, max_group_nid = -1;
1362 unsigned long max_faults = 0, max_group_faults = 0;
1363 unsigned long fault_types[2] = { 0, 0 };
1364 spinlock_t *group_lock = NULL;
1366 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1367 if (p->numa_scan_seq == seq)
1369 p->numa_scan_seq = seq;
1370 p->numa_scan_period_max = task_scan_max(p);
1372 /* If the task is part of a group prevent parallel updates to group stats */
1373 if (p->numa_group) {
1374 group_lock = &p->numa_group->lock;
1375 spin_lock(group_lock);
1378 /* Find the node with the highest number of faults */
1379 for_each_online_node(nid) {
1380 unsigned long faults = 0, group_faults = 0;
1383 for (priv = 0; priv < 2; priv++) {
1386 i = task_faults_idx(nid, priv);
1387 diff = -p->numa_faults[i];
1389 /* Decay existing window, copy faults since last scan */
1390 p->numa_faults[i] >>= 1;
1391 p->numa_faults[i] += p->numa_faults_buffer[i];
1392 fault_types[priv] += p->numa_faults_buffer[i];
1393 p->numa_faults_buffer[i] = 0;
1395 faults += p->numa_faults[i];
1396 diff += p->numa_faults[i];
1397 p->total_numa_faults += diff;
1398 if (p->numa_group) {
1399 /* safe because we can only change our own group */
1400 atomic_long_add(diff, &p->numa_group->faults[i]);
1401 atomic_long_add(diff, &p->numa_group->total_faults);
1402 group_faults += atomic_long_read(&p->numa_group->faults[i]);
1406 if (faults > max_faults) {
1407 max_faults = faults;
1411 if (group_faults > max_group_faults) {
1412 max_group_faults = group_faults;
1413 max_group_nid = nid;
1417 update_task_scan_period(p, fault_types[0], fault_types[1]);
1419 if (p->numa_group) {
1421 * If the preferred task and group nids are different,
1422 * iterate over the nodes again to find the best place.
1424 if (max_nid != max_group_nid) {
1425 unsigned long weight, max_weight = 0;
1427 for_each_online_node(nid) {
1428 weight = task_weight(p, nid) + group_weight(p, nid);
1429 if (weight > max_weight) {
1430 max_weight = weight;
1436 spin_unlock(group_lock);
1439 /* Preferred node as the node with the most faults */
1440 if (max_faults && max_nid != p->numa_preferred_nid) {
1441 /* Update the preferred nid and migrate task if possible */
1442 sched_setnuma(p, max_nid);
1443 numa_migrate_preferred(p);
1447 static inline int get_numa_group(struct numa_group *grp)
1449 return atomic_inc_not_zero(&grp->refcount);
1452 static inline void put_numa_group(struct numa_group *grp)
1454 if (atomic_dec_and_test(&grp->refcount))
1455 kfree_rcu(grp, rcu);
1458 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1464 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1467 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1470 struct numa_group *grp, *my_grp;
1471 struct task_struct *tsk;
1473 int cpu = cpupid_to_cpu(cpupid);
1476 if (unlikely(!p->numa_group)) {
1477 unsigned int size = sizeof(struct numa_group) +
1478 2*nr_node_ids*sizeof(atomic_long_t);
1480 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1484 atomic_set(&grp->refcount, 1);
1485 spin_lock_init(&grp->lock);
1486 INIT_LIST_HEAD(&grp->task_list);
1489 for (i = 0; i < 2*nr_node_ids; i++)
1490 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1492 atomic_long_set(&grp->total_faults, p->total_numa_faults);
1494 list_add(&p->numa_entry, &grp->task_list);
1496 rcu_assign_pointer(p->numa_group, grp);
1500 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1502 if (!cpupid_match_pid(tsk, cpupid))
1505 grp = rcu_dereference(tsk->numa_group);
1509 my_grp = p->numa_group;
1514 * Only join the other group if its bigger; if we're the bigger group,
1515 * the other task will join us.
1517 if (my_grp->nr_tasks > grp->nr_tasks)
1521 * Tie-break on the grp address.
1523 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1526 /* Always join threads in the same process. */
1527 if (tsk->mm == current->mm)
1530 /* Simple filter to avoid false positives due to PID collisions */
1531 if (flags & TNF_SHARED)
1534 /* Update priv based on whether false sharing was detected */
1537 if (join && !get_numa_group(grp))
1546 for (i = 0; i < 2*nr_node_ids; i++) {
1547 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1548 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1550 atomic_long_sub(p->total_numa_faults, &my_grp->total_faults);
1551 atomic_long_add(p->total_numa_faults, &grp->total_faults);
1553 double_lock(&my_grp->lock, &grp->lock);
1555 list_move(&p->numa_entry, &grp->task_list);
1559 spin_unlock(&my_grp->lock);
1560 spin_unlock(&grp->lock);
1562 rcu_assign_pointer(p->numa_group, grp);
1564 put_numa_group(my_grp);
1567 void task_numa_free(struct task_struct *p)
1569 struct numa_group *grp = p->numa_group;
1571 void *numa_faults = p->numa_faults;
1574 for (i = 0; i < 2*nr_node_ids; i++)
1575 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1577 atomic_long_sub(p->total_numa_faults, &grp->total_faults);
1579 spin_lock(&grp->lock);
1580 list_del(&p->numa_entry);
1582 spin_unlock(&grp->lock);
1583 rcu_assign_pointer(p->numa_group, NULL);
1584 put_numa_group(grp);
1587 p->numa_faults = NULL;
1588 p->numa_faults_buffer = NULL;
1593 * Got a PROT_NONE fault for a page on @node.
1595 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1597 struct task_struct *p = current;
1598 bool migrated = flags & TNF_MIGRATED;
1601 if (!numabalancing_enabled)
1604 /* for example, ksmd faulting in a user's mm */
1608 /* Do not worry about placement if exiting */
1609 if (p->state == TASK_DEAD)
1612 /* Allocate buffer to track faults on a per-node basis */
1613 if (unlikely(!p->numa_faults)) {
1614 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1616 /* numa_faults and numa_faults_buffer share the allocation */
1617 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1618 if (!p->numa_faults)
1621 BUG_ON(p->numa_faults_buffer);
1622 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1623 p->total_numa_faults = 0;
1624 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1628 * First accesses are treated as private, otherwise consider accesses
1629 * to be private if the accessing pid has not changed
1631 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1634 priv = cpupid_match_pid(p, last_cpupid);
1635 if (!priv && !(flags & TNF_NO_GROUP))
1636 task_numa_group(p, last_cpupid, flags, &priv);
1639 task_numa_placement(p);
1641 /* Retry task to preferred node migration if it previously failed */
1642 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1643 numa_migrate_preferred(p);
1646 p->numa_pages_migrated += pages;
1648 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1649 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1652 static void reset_ptenuma_scan(struct task_struct *p)
1654 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1655 p->mm->numa_scan_offset = 0;
1659 * The expensive part of numa migration is done from task_work context.
1660 * Triggered from task_tick_numa().
1662 void task_numa_work(struct callback_head *work)
1664 unsigned long migrate, next_scan, now = jiffies;
1665 struct task_struct *p = current;
1666 struct mm_struct *mm = p->mm;
1667 struct vm_area_struct *vma;
1668 unsigned long start, end;
1669 unsigned long nr_pte_updates = 0;
1672 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1674 work->next = work; /* protect against double add */
1676 * Who cares about NUMA placement when they're dying.
1678 * NOTE: make sure not to dereference p->mm before this check,
1679 * exit_task_work() happens _after_ exit_mm() so we could be called
1680 * without p->mm even though we still had it when we enqueued this
1683 if (p->flags & PF_EXITING)
1686 if (!mm->numa_next_scan) {
1687 mm->numa_next_scan = now +
1688 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1692 * Enforce maximal scan/migration frequency..
1694 migrate = mm->numa_next_scan;
1695 if (time_before(now, migrate))
1698 if (p->numa_scan_period == 0) {
1699 p->numa_scan_period_max = task_scan_max(p);
1700 p->numa_scan_period = task_scan_min(p);
1703 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1704 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1708 * Delay this task enough that another task of this mm will likely win
1709 * the next time around.
1711 p->node_stamp += 2 * TICK_NSEC;
1713 start = mm->numa_scan_offset;
1714 pages = sysctl_numa_balancing_scan_size;
1715 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1719 down_read(&mm->mmap_sem);
1720 vma = find_vma(mm, start);
1722 reset_ptenuma_scan(p);
1726 for (; vma; vma = vma->vm_next) {
1727 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1731 * Shared library pages mapped by multiple processes are not
1732 * migrated as it is expected they are cache replicated. Avoid
1733 * hinting faults in read-only file-backed mappings or the vdso
1734 * as migrating the pages will be of marginal benefit.
1737 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1741 start = max(start, vma->vm_start);
1742 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1743 end = min(end, vma->vm_end);
1744 nr_pte_updates += change_prot_numa(vma, start, end);
1747 * Scan sysctl_numa_balancing_scan_size but ensure that
1748 * at least one PTE is updated so that unused virtual
1749 * address space is quickly skipped.
1752 pages -= (end - start) >> PAGE_SHIFT;
1757 } while (end != vma->vm_end);
1762 * It is possible to reach the end of the VMA list but the last few
1763 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1764 * would find the !migratable VMA on the next scan but not reset the
1765 * scanner to the start so check it now.
1768 mm->numa_scan_offset = start;
1770 reset_ptenuma_scan(p);
1771 up_read(&mm->mmap_sem);
1775 * Drive the periodic memory faults..
1777 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1779 struct callback_head *work = &curr->numa_work;
1783 * We don't care about NUMA placement if we don't have memory.
1785 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1789 * Using runtime rather than walltime has the dual advantage that
1790 * we (mostly) drive the selection from busy threads and that the
1791 * task needs to have done some actual work before we bother with
1794 now = curr->se.sum_exec_runtime;
1795 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1797 if (now - curr->node_stamp > period) {
1798 if (!curr->node_stamp)
1799 curr->numa_scan_period = task_scan_min(curr);
1800 curr->node_stamp += period;
1802 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1803 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1804 task_work_add(curr, work, true);
1809 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1813 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1817 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1820 #endif /* CONFIG_NUMA_BALANCING */
1823 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1825 update_load_add(&cfs_rq->load, se->load.weight);
1826 if (!parent_entity(se))
1827 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1829 if (entity_is_task(se)) {
1830 struct rq *rq = rq_of(cfs_rq);
1832 account_numa_enqueue(rq, task_of(se));
1833 list_add(&se->group_node, &rq->cfs_tasks);
1836 cfs_rq->nr_running++;
1840 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1842 update_load_sub(&cfs_rq->load, se->load.weight);
1843 if (!parent_entity(se))
1844 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1845 if (entity_is_task(se)) {
1846 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1847 list_del_init(&se->group_node);
1849 cfs_rq->nr_running--;
1852 #ifdef CONFIG_FAIR_GROUP_SCHED
1854 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1859 * Use this CPU's actual weight instead of the last load_contribution
1860 * to gain a more accurate current total weight. See
1861 * update_cfs_rq_load_contribution().
1863 tg_weight = atomic_long_read(&tg->load_avg);
1864 tg_weight -= cfs_rq->tg_load_contrib;
1865 tg_weight += cfs_rq->load.weight;
1870 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1872 long tg_weight, load, shares;
1874 tg_weight = calc_tg_weight(tg, cfs_rq);
1875 load = cfs_rq->load.weight;
1877 shares = (tg->shares * load);
1879 shares /= tg_weight;
1881 if (shares < MIN_SHARES)
1882 shares = MIN_SHARES;
1883 if (shares > tg->shares)
1884 shares = tg->shares;
1888 # else /* CONFIG_SMP */
1889 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1893 # endif /* CONFIG_SMP */
1894 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1895 unsigned long weight)
1898 /* commit outstanding execution time */
1899 if (cfs_rq->curr == se)
1900 update_curr(cfs_rq);
1901 account_entity_dequeue(cfs_rq, se);
1904 update_load_set(&se->load, weight);
1907 account_entity_enqueue(cfs_rq, se);
1910 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1912 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1914 struct task_group *tg;
1915 struct sched_entity *se;
1919 se = tg->se[cpu_of(rq_of(cfs_rq))];
1920 if (!se || throttled_hierarchy(cfs_rq))
1923 if (likely(se->load.weight == tg->shares))
1926 shares = calc_cfs_shares(cfs_rq, tg);
1928 reweight_entity(cfs_rq_of(se), se, shares);
1930 #else /* CONFIG_FAIR_GROUP_SCHED */
1931 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1934 #endif /* CONFIG_FAIR_GROUP_SCHED */
1938 * We choose a half-life close to 1 scheduling period.
1939 * Note: The tables below are dependent on this value.
1941 #define LOAD_AVG_PERIOD 32
1942 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1943 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1945 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1946 static const u32 runnable_avg_yN_inv[] = {
1947 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1948 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1949 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1950 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1951 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1952 0x85aac367, 0x82cd8698,
1956 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1957 * over-estimates when re-combining.
1959 static const u32 runnable_avg_yN_sum[] = {
1960 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1961 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1962 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1967 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1969 static __always_inline u64 decay_load(u64 val, u64 n)
1971 unsigned int local_n;
1975 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1978 /* after bounds checking we can collapse to 32-bit */
1982 * As y^PERIOD = 1/2, we can combine
1983 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1984 * With a look-up table which covers k^n (n<PERIOD)
1986 * To achieve constant time decay_load.
1988 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1989 val >>= local_n / LOAD_AVG_PERIOD;
1990 local_n %= LOAD_AVG_PERIOD;
1993 val *= runnable_avg_yN_inv[local_n];
1994 /* We don't use SRR here since we always want to round down. */
1999 * For updates fully spanning n periods, the contribution to runnable
2000 * average will be: \Sum 1024*y^n
2002 * We can compute this reasonably efficiently by combining:
2003 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2005 static u32 __compute_runnable_contrib(u64 n)
2009 if (likely(n <= LOAD_AVG_PERIOD))
2010 return runnable_avg_yN_sum[n];
2011 else if (unlikely(n >= LOAD_AVG_MAX_N))
2012 return LOAD_AVG_MAX;
2014 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2016 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2017 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2019 n -= LOAD_AVG_PERIOD;
2020 } while (n > LOAD_AVG_PERIOD);
2022 contrib = decay_load(contrib, n);
2023 return contrib + runnable_avg_yN_sum[n];
2027 * We can represent the historical contribution to runnable average as the
2028 * coefficients of a geometric series. To do this we sub-divide our runnable
2029 * history into segments of approximately 1ms (1024us); label the segment that
2030 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2032 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2034 * (now) (~1ms ago) (~2ms ago)
2036 * Let u_i denote the fraction of p_i that the entity was runnable.
2038 * We then designate the fractions u_i as our co-efficients, yielding the
2039 * following representation of historical load:
2040 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2042 * We choose y based on the with of a reasonably scheduling period, fixing:
2045 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2046 * approximately half as much as the contribution to load within the last ms
2049 * When a period "rolls over" and we have new u_0`, multiplying the previous
2050 * sum again by y is sufficient to update:
2051 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2052 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2054 static __always_inline int __update_entity_runnable_avg(u64 now,
2055 struct sched_avg *sa,
2059 u32 runnable_contrib;
2060 int delta_w, decayed = 0;
2062 delta = now - sa->last_runnable_update;
2064 * This should only happen when time goes backwards, which it
2065 * unfortunately does during sched clock init when we swap over to TSC.
2067 if ((s64)delta < 0) {
2068 sa->last_runnable_update = now;
2073 * Use 1024ns as the unit of measurement since it's a reasonable
2074 * approximation of 1us and fast to compute.
2079 sa->last_runnable_update = now;
2081 /* delta_w is the amount already accumulated against our next period */
2082 delta_w = sa->runnable_avg_period % 1024;
2083 if (delta + delta_w >= 1024) {
2084 /* period roll-over */
2088 * Now that we know we're crossing a period boundary, figure
2089 * out how much from delta we need to complete the current
2090 * period and accrue it.
2092 delta_w = 1024 - delta_w;
2094 sa->runnable_avg_sum += delta_w;
2095 sa->runnable_avg_period += delta_w;
2099 /* Figure out how many additional periods this update spans */
2100 periods = delta / 1024;
2103 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2105 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2108 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2109 runnable_contrib = __compute_runnable_contrib(periods);
2111 sa->runnable_avg_sum += runnable_contrib;
2112 sa->runnable_avg_period += runnable_contrib;
2115 /* Remainder of delta accrued against u_0` */
2117 sa->runnable_avg_sum += delta;
2118 sa->runnable_avg_period += delta;
2123 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2124 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2126 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2127 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2129 decays -= se->avg.decay_count;
2133 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2134 se->avg.decay_count = 0;
2139 #ifdef CONFIG_FAIR_GROUP_SCHED
2140 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2143 struct task_group *tg = cfs_rq->tg;
2146 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2147 tg_contrib -= cfs_rq->tg_load_contrib;
2149 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2150 atomic_long_add(tg_contrib, &tg->load_avg);
2151 cfs_rq->tg_load_contrib += tg_contrib;
2156 * Aggregate cfs_rq runnable averages into an equivalent task_group
2157 * representation for computing load contributions.
2159 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2160 struct cfs_rq *cfs_rq)
2162 struct task_group *tg = cfs_rq->tg;
2165 /* The fraction of a cpu used by this cfs_rq */
2166 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2167 sa->runnable_avg_period + 1);
2168 contrib -= cfs_rq->tg_runnable_contrib;
2170 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2171 atomic_add(contrib, &tg->runnable_avg);
2172 cfs_rq->tg_runnable_contrib += contrib;
2176 static inline void __update_group_entity_contrib(struct sched_entity *se)
2178 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2179 struct task_group *tg = cfs_rq->tg;
2184 contrib = cfs_rq->tg_load_contrib * tg->shares;
2185 se->avg.load_avg_contrib = div_u64(contrib,
2186 atomic_long_read(&tg->load_avg) + 1);
2189 * For group entities we need to compute a correction term in the case
2190 * that they are consuming <1 cpu so that we would contribute the same
2191 * load as a task of equal weight.
2193 * Explicitly co-ordinating this measurement would be expensive, but
2194 * fortunately the sum of each cpus contribution forms a usable
2195 * lower-bound on the true value.
2197 * Consider the aggregate of 2 contributions. Either they are disjoint
2198 * (and the sum represents true value) or they are disjoint and we are
2199 * understating by the aggregate of their overlap.
2201 * Extending this to N cpus, for a given overlap, the maximum amount we
2202 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2203 * cpus that overlap for this interval and w_i is the interval width.
2205 * On a small machine; the first term is well-bounded which bounds the
2206 * total error since w_i is a subset of the period. Whereas on a
2207 * larger machine, while this first term can be larger, if w_i is the
2208 * of consequential size guaranteed to see n_i*w_i quickly converge to
2209 * our upper bound of 1-cpu.
2211 runnable_avg = atomic_read(&tg->runnable_avg);
2212 if (runnable_avg < NICE_0_LOAD) {
2213 se->avg.load_avg_contrib *= runnable_avg;
2214 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2218 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2219 int force_update) {}
2220 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2221 struct cfs_rq *cfs_rq) {}
2222 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2225 static inline void __update_task_entity_contrib(struct sched_entity *se)
2229 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2230 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2231 contrib /= (se->avg.runnable_avg_period + 1);
2232 se->avg.load_avg_contrib = scale_load(contrib);
2235 /* Compute the current contribution to load_avg by se, return any delta */
2236 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2238 long old_contrib = se->avg.load_avg_contrib;
2240 if (entity_is_task(se)) {
2241 __update_task_entity_contrib(se);
2243 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2244 __update_group_entity_contrib(se);
2247 return se->avg.load_avg_contrib - old_contrib;
2250 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2253 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2254 cfs_rq->blocked_load_avg -= load_contrib;
2256 cfs_rq->blocked_load_avg = 0;
2259 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2261 /* Update a sched_entity's runnable average */
2262 static inline void update_entity_load_avg(struct sched_entity *se,
2265 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2270 * For a group entity we need to use their owned cfs_rq_clock_task() in
2271 * case they are the parent of a throttled hierarchy.
2273 if (entity_is_task(se))
2274 now = cfs_rq_clock_task(cfs_rq);
2276 now = cfs_rq_clock_task(group_cfs_rq(se));
2278 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2281 contrib_delta = __update_entity_load_avg_contrib(se);
2287 cfs_rq->runnable_load_avg += contrib_delta;
2289 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2293 * Decay the load contributed by all blocked children and account this so that
2294 * their contribution may appropriately discounted when they wake up.
2296 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2298 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2301 decays = now - cfs_rq->last_decay;
2302 if (!decays && !force_update)
2305 if (atomic_long_read(&cfs_rq->removed_load)) {
2306 unsigned long removed_load;
2307 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2308 subtract_blocked_load_contrib(cfs_rq, removed_load);
2312 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2314 atomic64_add(decays, &cfs_rq->decay_counter);
2315 cfs_rq->last_decay = now;
2318 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2321 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2323 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2324 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2327 /* Add the load generated by se into cfs_rq's child load-average */
2328 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2329 struct sched_entity *se,
2333 * We track migrations using entity decay_count <= 0, on a wake-up
2334 * migration we use a negative decay count to track the remote decays
2335 * accumulated while sleeping.
2337 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2338 * are seen by enqueue_entity_load_avg() as a migration with an already
2339 * constructed load_avg_contrib.
2341 if (unlikely(se->avg.decay_count <= 0)) {
2342 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2343 if (se->avg.decay_count) {
2345 * In a wake-up migration we have to approximate the
2346 * time sleeping. This is because we can't synchronize
2347 * clock_task between the two cpus, and it is not
2348 * guaranteed to be read-safe. Instead, we can
2349 * approximate this using our carried decays, which are
2350 * explicitly atomically readable.
2352 se->avg.last_runnable_update -= (-se->avg.decay_count)
2354 update_entity_load_avg(se, 0);
2355 /* Indicate that we're now synchronized and on-rq */
2356 se->avg.decay_count = 0;
2361 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2362 * would have made count negative); we must be careful to avoid
2363 * double-accounting blocked time after synchronizing decays.
2365 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2369 /* migrated tasks did not contribute to our blocked load */
2371 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2372 update_entity_load_avg(se, 0);
2375 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2376 /* we force update consideration on load-balancer moves */
2377 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2381 * Remove se's load from this cfs_rq child load-average, if the entity is
2382 * transitioning to a blocked state we track its projected decay using
2385 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2386 struct sched_entity *se,
2389 update_entity_load_avg(se, 1);
2390 /* we force update consideration on load-balancer moves */
2391 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2393 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2395 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2396 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2397 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2401 * Update the rq's load with the elapsed running time before entering
2402 * idle. if the last scheduled task is not a CFS task, idle_enter will
2403 * be the only way to update the runnable statistic.
2405 void idle_enter_fair(struct rq *this_rq)
2407 update_rq_runnable_avg(this_rq, 1);
2411 * Update the rq's load with the elapsed idle time before a task is
2412 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2413 * be the only way to update the runnable statistic.
2415 void idle_exit_fair(struct rq *this_rq)
2417 update_rq_runnable_avg(this_rq, 0);
2421 static inline void update_entity_load_avg(struct sched_entity *se,
2422 int update_cfs_rq) {}
2423 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2424 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2425 struct sched_entity *se,
2427 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2428 struct sched_entity *se,
2430 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2431 int force_update) {}
2434 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2436 #ifdef CONFIG_SCHEDSTATS
2437 struct task_struct *tsk = NULL;
2439 if (entity_is_task(se))
2442 if (se->statistics.sleep_start) {
2443 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2448 if (unlikely(delta > se->statistics.sleep_max))
2449 se->statistics.sleep_max = delta;
2451 se->statistics.sleep_start = 0;
2452 se->statistics.sum_sleep_runtime += delta;
2455 account_scheduler_latency(tsk, delta >> 10, 1);
2456 trace_sched_stat_sleep(tsk, delta);
2459 if (se->statistics.block_start) {
2460 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2465 if (unlikely(delta > se->statistics.block_max))
2466 se->statistics.block_max = delta;
2468 se->statistics.block_start = 0;
2469 se->statistics.sum_sleep_runtime += delta;
2472 if (tsk->in_iowait) {
2473 se->statistics.iowait_sum += delta;
2474 se->statistics.iowait_count++;
2475 trace_sched_stat_iowait(tsk, delta);
2478 trace_sched_stat_blocked(tsk, delta);
2481 * Blocking time is in units of nanosecs, so shift by
2482 * 20 to get a milliseconds-range estimation of the
2483 * amount of time that the task spent sleeping:
2485 if (unlikely(prof_on == SLEEP_PROFILING)) {
2486 profile_hits(SLEEP_PROFILING,
2487 (void *)get_wchan(tsk),
2490 account_scheduler_latency(tsk, delta >> 10, 0);
2496 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2498 #ifdef CONFIG_SCHED_DEBUG
2499 s64 d = se->vruntime - cfs_rq->min_vruntime;
2504 if (d > 3*sysctl_sched_latency)
2505 schedstat_inc(cfs_rq, nr_spread_over);
2510 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2512 u64 vruntime = cfs_rq->min_vruntime;
2515 * The 'current' period is already promised to the current tasks,
2516 * however the extra weight of the new task will slow them down a
2517 * little, place the new task so that it fits in the slot that
2518 * stays open at the end.
2520 if (initial && sched_feat(START_DEBIT))
2521 vruntime += sched_vslice(cfs_rq, se);
2523 /* sleeps up to a single latency don't count. */
2525 unsigned long thresh = sysctl_sched_latency;
2528 * Halve their sleep time's effect, to allow
2529 * for a gentler effect of sleepers:
2531 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2537 /* ensure we never gain time by being placed backwards. */
2538 se->vruntime = max_vruntime(se->vruntime, vruntime);
2541 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2544 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2547 * Update the normalized vruntime before updating min_vruntime
2548 * through calling update_curr().
2550 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2551 se->vruntime += cfs_rq->min_vruntime;
2554 * Update run-time statistics of the 'current'.
2556 update_curr(cfs_rq);
2557 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2558 account_entity_enqueue(cfs_rq, se);
2559 update_cfs_shares(cfs_rq);
2561 if (flags & ENQUEUE_WAKEUP) {
2562 place_entity(cfs_rq, se, 0);
2563 enqueue_sleeper(cfs_rq, se);
2566 update_stats_enqueue(cfs_rq, se);
2567 check_spread(cfs_rq, se);
2568 if (se != cfs_rq->curr)
2569 __enqueue_entity(cfs_rq, se);
2572 if (cfs_rq->nr_running == 1) {
2573 list_add_leaf_cfs_rq(cfs_rq);
2574 check_enqueue_throttle(cfs_rq);
2578 static void __clear_buddies_last(struct sched_entity *se)
2580 for_each_sched_entity(se) {
2581 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2582 if (cfs_rq->last == se)
2583 cfs_rq->last = NULL;
2589 static void __clear_buddies_next(struct sched_entity *se)
2591 for_each_sched_entity(se) {
2592 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2593 if (cfs_rq->next == se)
2594 cfs_rq->next = NULL;
2600 static void __clear_buddies_skip(struct sched_entity *se)
2602 for_each_sched_entity(se) {
2603 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2604 if (cfs_rq->skip == se)
2605 cfs_rq->skip = NULL;
2611 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2613 if (cfs_rq->last == se)
2614 __clear_buddies_last(se);
2616 if (cfs_rq->next == se)
2617 __clear_buddies_next(se);
2619 if (cfs_rq->skip == se)
2620 __clear_buddies_skip(se);
2623 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2626 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2629 * Update run-time statistics of the 'current'.
2631 update_curr(cfs_rq);
2632 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2634 update_stats_dequeue(cfs_rq, se);
2635 if (flags & DEQUEUE_SLEEP) {
2636 #ifdef CONFIG_SCHEDSTATS
2637 if (entity_is_task(se)) {
2638 struct task_struct *tsk = task_of(se);
2640 if (tsk->state & TASK_INTERRUPTIBLE)
2641 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2642 if (tsk->state & TASK_UNINTERRUPTIBLE)
2643 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2648 clear_buddies(cfs_rq, se);
2650 if (se != cfs_rq->curr)
2651 __dequeue_entity(cfs_rq, se);
2653 account_entity_dequeue(cfs_rq, se);
2656 * Normalize the entity after updating the min_vruntime because the
2657 * update can refer to the ->curr item and we need to reflect this
2658 * movement in our normalized position.
2660 if (!(flags & DEQUEUE_SLEEP))
2661 se->vruntime -= cfs_rq->min_vruntime;
2663 /* return excess runtime on last dequeue */
2664 return_cfs_rq_runtime(cfs_rq);
2666 update_min_vruntime(cfs_rq);
2667 update_cfs_shares(cfs_rq);
2671 * Preempt the current task with a newly woken task if needed:
2674 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2676 unsigned long ideal_runtime, delta_exec;
2677 struct sched_entity *se;
2680 ideal_runtime = sched_slice(cfs_rq, curr);
2681 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2682 if (delta_exec > ideal_runtime) {
2683 resched_task(rq_of(cfs_rq)->curr);
2685 * The current task ran long enough, ensure it doesn't get
2686 * re-elected due to buddy favours.
2688 clear_buddies(cfs_rq, curr);
2693 * Ensure that a task that missed wakeup preemption by a
2694 * narrow margin doesn't have to wait for a full slice.
2695 * This also mitigates buddy induced latencies under load.
2697 if (delta_exec < sysctl_sched_min_granularity)
2700 se = __pick_first_entity(cfs_rq);
2701 delta = curr->vruntime - se->vruntime;
2706 if (delta > ideal_runtime)
2707 resched_task(rq_of(cfs_rq)->curr);
2711 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2713 /* 'current' is not kept within the tree. */
2716 * Any task has to be enqueued before it get to execute on
2717 * a CPU. So account for the time it spent waiting on the
2720 update_stats_wait_end(cfs_rq, se);
2721 __dequeue_entity(cfs_rq, se);
2724 update_stats_curr_start(cfs_rq, se);
2726 #ifdef CONFIG_SCHEDSTATS
2728 * Track our maximum slice length, if the CPU's load is at
2729 * least twice that of our own weight (i.e. dont track it
2730 * when there are only lesser-weight tasks around):
2732 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2733 se->statistics.slice_max = max(se->statistics.slice_max,
2734 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2737 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2741 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2744 * Pick the next process, keeping these things in mind, in this order:
2745 * 1) keep things fair between processes/task groups
2746 * 2) pick the "next" process, since someone really wants that to run
2747 * 3) pick the "last" process, for cache locality
2748 * 4) do not run the "skip" process, if something else is available
2750 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2752 struct sched_entity *se = __pick_first_entity(cfs_rq);
2753 struct sched_entity *left = se;
2756 * Avoid running the skip buddy, if running something else can
2757 * be done without getting too unfair.
2759 if (cfs_rq->skip == se) {
2760 struct sched_entity *second = __pick_next_entity(se);
2761 if (second && wakeup_preempt_entity(second, left) < 1)
2766 * Prefer last buddy, try to return the CPU to a preempted task.
2768 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2772 * Someone really wants this to run. If it's not unfair, run it.
2774 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2777 clear_buddies(cfs_rq, se);
2782 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2784 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2787 * If still on the runqueue then deactivate_task()
2788 * was not called and update_curr() has to be done:
2791 update_curr(cfs_rq);
2793 /* throttle cfs_rqs exceeding runtime */
2794 check_cfs_rq_runtime(cfs_rq);
2796 check_spread(cfs_rq, prev);
2798 update_stats_wait_start(cfs_rq, prev);
2799 /* Put 'current' back into the tree. */
2800 __enqueue_entity(cfs_rq, prev);
2801 /* in !on_rq case, update occurred at dequeue */
2802 update_entity_load_avg(prev, 1);
2804 cfs_rq->curr = NULL;
2808 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2811 * Update run-time statistics of the 'current'.
2813 update_curr(cfs_rq);
2816 * Ensure that runnable average is periodically updated.
2818 update_entity_load_avg(curr, 1);
2819 update_cfs_rq_blocked_load(cfs_rq, 1);
2820 update_cfs_shares(cfs_rq);
2822 #ifdef CONFIG_SCHED_HRTICK
2824 * queued ticks are scheduled to match the slice, so don't bother
2825 * validating it and just reschedule.
2828 resched_task(rq_of(cfs_rq)->curr);
2832 * don't let the period tick interfere with the hrtick preemption
2834 if (!sched_feat(DOUBLE_TICK) &&
2835 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2839 if (cfs_rq->nr_running > 1)
2840 check_preempt_tick(cfs_rq, curr);
2844 /**************************************************
2845 * CFS bandwidth control machinery
2848 #ifdef CONFIG_CFS_BANDWIDTH
2850 #ifdef HAVE_JUMP_LABEL
2851 static struct static_key __cfs_bandwidth_used;
2853 static inline bool cfs_bandwidth_used(void)
2855 return static_key_false(&__cfs_bandwidth_used);
2858 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2860 /* only need to count groups transitioning between enabled/!enabled */
2861 if (enabled && !was_enabled)
2862 static_key_slow_inc(&__cfs_bandwidth_used);
2863 else if (!enabled && was_enabled)
2864 static_key_slow_dec(&__cfs_bandwidth_used);
2866 #else /* HAVE_JUMP_LABEL */
2867 static bool cfs_bandwidth_used(void)
2872 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2873 #endif /* HAVE_JUMP_LABEL */
2876 * default period for cfs group bandwidth.
2877 * default: 0.1s, units: nanoseconds
2879 static inline u64 default_cfs_period(void)
2881 return 100000000ULL;
2884 static inline u64 sched_cfs_bandwidth_slice(void)
2886 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2890 * Replenish runtime according to assigned quota and update expiration time.
2891 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2892 * additional synchronization around rq->lock.
2894 * requires cfs_b->lock
2896 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2900 if (cfs_b->quota == RUNTIME_INF)
2903 now = sched_clock_cpu(smp_processor_id());
2904 cfs_b->runtime = cfs_b->quota;
2905 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2908 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2910 return &tg->cfs_bandwidth;
2913 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2914 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2916 if (unlikely(cfs_rq->throttle_count))
2917 return cfs_rq->throttled_clock_task;
2919 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2922 /* returns 0 on failure to allocate runtime */
2923 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2925 struct task_group *tg = cfs_rq->tg;
2926 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2927 u64 amount = 0, min_amount, expires;
2929 /* note: this is a positive sum as runtime_remaining <= 0 */
2930 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2932 raw_spin_lock(&cfs_b->lock);
2933 if (cfs_b->quota == RUNTIME_INF)
2934 amount = min_amount;
2937 * If the bandwidth pool has become inactive, then at least one
2938 * period must have elapsed since the last consumption.
2939 * Refresh the global state and ensure bandwidth timer becomes
2942 if (!cfs_b->timer_active) {
2943 __refill_cfs_bandwidth_runtime(cfs_b);
2944 __start_cfs_bandwidth(cfs_b);
2947 if (cfs_b->runtime > 0) {
2948 amount = min(cfs_b->runtime, min_amount);
2949 cfs_b->runtime -= amount;
2953 expires = cfs_b->runtime_expires;
2954 raw_spin_unlock(&cfs_b->lock);
2956 cfs_rq->runtime_remaining += amount;
2958 * we may have advanced our local expiration to account for allowed
2959 * spread between our sched_clock and the one on which runtime was
2962 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2963 cfs_rq->runtime_expires = expires;
2965 return cfs_rq->runtime_remaining > 0;
2969 * Note: This depends on the synchronization provided by sched_clock and the
2970 * fact that rq->clock snapshots this value.
2972 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2974 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2976 /* if the deadline is ahead of our clock, nothing to do */
2977 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2980 if (cfs_rq->runtime_remaining < 0)
2984 * If the local deadline has passed we have to consider the
2985 * possibility that our sched_clock is 'fast' and the global deadline
2986 * has not truly expired.
2988 * Fortunately we can check determine whether this the case by checking
2989 * whether the global deadline has advanced.
2992 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2993 /* extend local deadline, drift is bounded above by 2 ticks */
2994 cfs_rq->runtime_expires += TICK_NSEC;
2996 /* global deadline is ahead, expiration has passed */
2997 cfs_rq->runtime_remaining = 0;
3001 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3002 unsigned long delta_exec)
3004 /* dock delta_exec before expiring quota (as it could span periods) */
3005 cfs_rq->runtime_remaining -= delta_exec;
3006 expire_cfs_rq_runtime(cfs_rq);
3008 if (likely(cfs_rq->runtime_remaining > 0))
3012 * if we're unable to extend our runtime we resched so that the active
3013 * hierarchy can be throttled
3015 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3016 resched_task(rq_of(cfs_rq)->curr);
3019 static __always_inline
3020 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
3022 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3025 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3028 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3030 return cfs_bandwidth_used() && cfs_rq->throttled;
3033 /* check whether cfs_rq, or any parent, is throttled */
3034 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3036 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3040 * Ensure that neither of the group entities corresponding to src_cpu or
3041 * dest_cpu are members of a throttled hierarchy when performing group
3042 * load-balance operations.
3044 static inline int throttled_lb_pair(struct task_group *tg,
3045 int src_cpu, int dest_cpu)
3047 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3049 src_cfs_rq = tg->cfs_rq[src_cpu];
3050 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3052 return throttled_hierarchy(src_cfs_rq) ||
3053 throttled_hierarchy(dest_cfs_rq);
3056 /* updated child weight may affect parent so we have to do this bottom up */
3057 static int tg_unthrottle_up(struct task_group *tg, void *data)
3059 struct rq *rq = data;
3060 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3062 cfs_rq->throttle_count--;
3064 if (!cfs_rq->throttle_count) {
3065 /* adjust cfs_rq_clock_task() */
3066 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3067 cfs_rq->throttled_clock_task;
3074 static int tg_throttle_down(struct task_group *tg, void *data)
3076 struct rq *rq = data;
3077 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3079 /* group is entering throttled state, stop time */
3080 if (!cfs_rq->throttle_count)
3081 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3082 cfs_rq->throttle_count++;
3087 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3089 struct rq *rq = rq_of(cfs_rq);
3090 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3091 struct sched_entity *se;
3092 long task_delta, dequeue = 1;
3094 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3096 /* freeze hierarchy runnable averages while throttled */
3098 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3101 task_delta = cfs_rq->h_nr_running;
3102 for_each_sched_entity(se) {
3103 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3104 /* throttled entity or throttle-on-deactivate */
3109 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3110 qcfs_rq->h_nr_running -= task_delta;
3112 if (qcfs_rq->load.weight)
3117 rq->nr_running -= task_delta;
3119 cfs_rq->throttled = 1;
3120 cfs_rq->throttled_clock = rq_clock(rq);
3121 raw_spin_lock(&cfs_b->lock);
3122 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3123 raw_spin_unlock(&cfs_b->lock);
3126 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3128 struct rq *rq = rq_of(cfs_rq);
3129 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3130 struct sched_entity *se;
3134 se = cfs_rq->tg->se[cpu_of(rq)];
3136 cfs_rq->throttled = 0;
3138 update_rq_clock(rq);
3140 raw_spin_lock(&cfs_b->lock);
3141 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3142 list_del_rcu(&cfs_rq->throttled_list);
3143 raw_spin_unlock(&cfs_b->lock);
3145 /* update hierarchical throttle state */
3146 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3148 if (!cfs_rq->load.weight)
3151 task_delta = cfs_rq->h_nr_running;
3152 for_each_sched_entity(se) {
3156 cfs_rq = cfs_rq_of(se);
3158 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3159 cfs_rq->h_nr_running += task_delta;
3161 if (cfs_rq_throttled(cfs_rq))
3166 rq->nr_running += task_delta;
3168 /* determine whether we need to wake up potentially idle cpu */
3169 if (rq->curr == rq->idle && rq->cfs.nr_running)
3170 resched_task(rq->curr);
3173 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3174 u64 remaining, u64 expires)
3176 struct cfs_rq *cfs_rq;
3177 u64 runtime = remaining;
3180 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3182 struct rq *rq = rq_of(cfs_rq);
3184 raw_spin_lock(&rq->lock);
3185 if (!cfs_rq_throttled(cfs_rq))
3188 runtime = -cfs_rq->runtime_remaining + 1;
3189 if (runtime > remaining)
3190 runtime = remaining;
3191 remaining -= runtime;
3193 cfs_rq->runtime_remaining += runtime;
3194 cfs_rq->runtime_expires = expires;
3196 /* we check whether we're throttled above */
3197 if (cfs_rq->runtime_remaining > 0)
3198 unthrottle_cfs_rq(cfs_rq);
3201 raw_spin_unlock(&rq->lock);
3212 * Responsible for refilling a task_group's bandwidth and unthrottling its
3213 * cfs_rqs as appropriate. If there has been no activity within the last
3214 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3215 * used to track this state.
3217 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3219 u64 runtime, runtime_expires;
3220 int idle = 1, throttled;
3222 raw_spin_lock(&cfs_b->lock);
3223 /* no need to continue the timer with no bandwidth constraint */
3224 if (cfs_b->quota == RUNTIME_INF)
3227 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3228 /* idle depends on !throttled (for the case of a large deficit) */
3229 idle = cfs_b->idle && !throttled;
3230 cfs_b->nr_periods += overrun;
3232 /* if we're going inactive then everything else can be deferred */
3236 __refill_cfs_bandwidth_runtime(cfs_b);
3239 /* mark as potentially idle for the upcoming period */
3244 /* account preceding periods in which throttling occurred */
3245 cfs_b->nr_throttled += overrun;
3248 * There are throttled entities so we must first use the new bandwidth
3249 * to unthrottle them before making it generally available. This
3250 * ensures that all existing debts will be paid before a new cfs_rq is
3253 runtime = cfs_b->runtime;
3254 runtime_expires = cfs_b->runtime_expires;
3258 * This check is repeated as we are holding onto the new bandwidth
3259 * while we unthrottle. This can potentially race with an unthrottled
3260 * group trying to acquire new bandwidth from the global pool.
3262 while (throttled && runtime > 0) {
3263 raw_spin_unlock(&cfs_b->lock);
3264 /* we can't nest cfs_b->lock while distributing bandwidth */
3265 runtime = distribute_cfs_runtime(cfs_b, runtime,
3267 raw_spin_lock(&cfs_b->lock);
3269 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3272 /* return (any) remaining runtime */
3273 cfs_b->runtime = runtime;
3275 * While we are ensured activity in the period following an
3276 * unthrottle, this also covers the case in which the new bandwidth is
3277 * insufficient to cover the existing bandwidth deficit. (Forcing the
3278 * timer to remain active while there are any throttled entities.)
3283 cfs_b->timer_active = 0;
3284 raw_spin_unlock(&cfs_b->lock);
3289 /* a cfs_rq won't donate quota below this amount */
3290 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3291 /* minimum remaining period time to redistribute slack quota */
3292 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3293 /* how long we wait to gather additional slack before distributing */
3294 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3296 /* are we near the end of the current quota period? */
3297 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3299 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3302 /* if the call-back is running a quota refresh is already occurring */
3303 if (hrtimer_callback_running(refresh_timer))
3306 /* is a quota refresh about to occur? */
3307 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3308 if (remaining < min_expire)
3314 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3316 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3318 /* if there's a quota refresh soon don't bother with slack */
3319 if (runtime_refresh_within(cfs_b, min_left))
3322 start_bandwidth_timer(&cfs_b->slack_timer,
3323 ns_to_ktime(cfs_bandwidth_slack_period));
3326 /* we know any runtime found here is valid as update_curr() precedes return */
3327 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3329 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3330 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3332 if (slack_runtime <= 0)
3335 raw_spin_lock(&cfs_b->lock);
3336 if (cfs_b->quota != RUNTIME_INF &&
3337 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3338 cfs_b->runtime += slack_runtime;
3340 /* we are under rq->lock, defer unthrottling using a timer */
3341 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3342 !list_empty(&cfs_b->throttled_cfs_rq))
3343 start_cfs_slack_bandwidth(cfs_b);
3345 raw_spin_unlock(&cfs_b->lock);
3347 /* even if it's not valid for return we don't want to try again */
3348 cfs_rq->runtime_remaining -= slack_runtime;
3351 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3353 if (!cfs_bandwidth_used())
3356 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3359 __return_cfs_rq_runtime(cfs_rq);
3363 * This is done with a timer (instead of inline with bandwidth return) since
3364 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3366 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3368 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3371 /* confirm we're still not at a refresh boundary */
3372 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3375 raw_spin_lock(&cfs_b->lock);
3376 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3377 runtime = cfs_b->runtime;
3380 expires = cfs_b->runtime_expires;
3381 raw_spin_unlock(&cfs_b->lock);
3386 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3388 raw_spin_lock(&cfs_b->lock);
3389 if (expires == cfs_b->runtime_expires)
3390 cfs_b->runtime = runtime;
3391 raw_spin_unlock(&cfs_b->lock);
3395 * When a group wakes up we want to make sure that its quota is not already
3396 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3397 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3399 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3401 if (!cfs_bandwidth_used())
3404 /* an active group must be handled by the update_curr()->put() path */
3405 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3408 /* ensure the group is not already throttled */
3409 if (cfs_rq_throttled(cfs_rq))
3412 /* update runtime allocation */
3413 account_cfs_rq_runtime(cfs_rq, 0);
3414 if (cfs_rq->runtime_remaining <= 0)
3415 throttle_cfs_rq(cfs_rq);
3418 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3419 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3421 if (!cfs_bandwidth_used())
3424 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3428 * it's possible for a throttled entity to be forced into a running
3429 * state (e.g. set_curr_task), in this case we're finished.
3431 if (cfs_rq_throttled(cfs_rq))
3434 throttle_cfs_rq(cfs_rq);
3437 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3439 struct cfs_bandwidth *cfs_b =
3440 container_of(timer, struct cfs_bandwidth, slack_timer);
3441 do_sched_cfs_slack_timer(cfs_b);
3443 return HRTIMER_NORESTART;
3446 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3448 struct cfs_bandwidth *cfs_b =
3449 container_of(timer, struct cfs_bandwidth, period_timer);
3455 now = hrtimer_cb_get_time(timer);
3456 overrun = hrtimer_forward(timer, now, cfs_b->period);
3461 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3464 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3467 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3469 raw_spin_lock_init(&cfs_b->lock);
3471 cfs_b->quota = RUNTIME_INF;
3472 cfs_b->period = ns_to_ktime(default_cfs_period());
3474 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3475 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3476 cfs_b->period_timer.function = sched_cfs_period_timer;
3477 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3478 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3481 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3483 cfs_rq->runtime_enabled = 0;
3484 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3487 /* requires cfs_b->lock, may release to reprogram timer */
3488 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3491 * The timer may be active because we're trying to set a new bandwidth
3492 * period or because we're racing with the tear-down path
3493 * (timer_active==0 becomes visible before the hrtimer call-back
3494 * terminates). In either case we ensure that it's re-programmed
3496 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3497 raw_spin_unlock(&cfs_b->lock);
3498 /* ensure cfs_b->lock is available while we wait */
3499 hrtimer_cancel(&cfs_b->period_timer);
3501 raw_spin_lock(&cfs_b->lock);
3502 /* if someone else restarted the timer then we're done */
3503 if (cfs_b->timer_active)
3507 cfs_b->timer_active = 1;
3508 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3511 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3513 hrtimer_cancel(&cfs_b->period_timer);
3514 hrtimer_cancel(&cfs_b->slack_timer);
3517 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3519 struct cfs_rq *cfs_rq;
3521 for_each_leaf_cfs_rq(rq, cfs_rq) {
3522 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3524 if (!cfs_rq->runtime_enabled)
3528 * clock_task is not advancing so we just need to make sure
3529 * there's some valid quota amount
3531 cfs_rq->runtime_remaining = cfs_b->quota;
3532 if (cfs_rq_throttled(cfs_rq))
3533 unthrottle_cfs_rq(cfs_rq);
3537 #else /* CONFIG_CFS_BANDWIDTH */
3538 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3540 return rq_clock_task(rq_of(cfs_rq));
3543 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3544 unsigned long delta_exec) {}
3545 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3546 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3547 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3549 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3554 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3559 static inline int throttled_lb_pair(struct task_group *tg,
3560 int src_cpu, int dest_cpu)
3565 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3567 #ifdef CONFIG_FAIR_GROUP_SCHED
3568 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3571 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3575 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3576 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3578 #endif /* CONFIG_CFS_BANDWIDTH */
3580 /**************************************************
3581 * CFS operations on tasks:
3584 #ifdef CONFIG_SCHED_HRTICK
3585 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3587 struct sched_entity *se = &p->se;
3588 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3590 WARN_ON(task_rq(p) != rq);
3592 if (cfs_rq->nr_running > 1) {
3593 u64 slice = sched_slice(cfs_rq, se);
3594 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3595 s64 delta = slice - ran;
3604 * Don't schedule slices shorter than 10000ns, that just
3605 * doesn't make sense. Rely on vruntime for fairness.
3608 delta = max_t(s64, 10000LL, delta);
3610 hrtick_start(rq, delta);
3615 * called from enqueue/dequeue and updates the hrtick when the
3616 * current task is from our class and nr_running is low enough
3619 static void hrtick_update(struct rq *rq)
3621 struct task_struct *curr = rq->curr;
3623 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3626 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3627 hrtick_start_fair(rq, curr);
3629 #else /* !CONFIG_SCHED_HRTICK */
3631 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3635 static inline void hrtick_update(struct rq *rq)
3641 * The enqueue_task method is called before nr_running is
3642 * increased. Here we update the fair scheduling stats and
3643 * then put the task into the rbtree:
3646 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3648 struct cfs_rq *cfs_rq;
3649 struct sched_entity *se = &p->se;
3651 for_each_sched_entity(se) {
3654 cfs_rq = cfs_rq_of(se);
3655 enqueue_entity(cfs_rq, se, flags);
3658 * end evaluation on encountering a throttled cfs_rq
3660 * note: in the case of encountering a throttled cfs_rq we will
3661 * post the final h_nr_running increment below.
3663 if (cfs_rq_throttled(cfs_rq))
3665 cfs_rq->h_nr_running++;
3667 flags = ENQUEUE_WAKEUP;
3670 for_each_sched_entity(se) {
3671 cfs_rq = cfs_rq_of(se);
3672 cfs_rq->h_nr_running++;
3674 if (cfs_rq_throttled(cfs_rq))
3677 update_cfs_shares(cfs_rq);
3678 update_entity_load_avg(se, 1);
3682 update_rq_runnable_avg(rq, rq->nr_running);
3688 static void set_next_buddy(struct sched_entity *se);
3691 * The dequeue_task method is called before nr_running is
3692 * decreased. We remove the task from the rbtree and
3693 * update the fair scheduling stats:
3695 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3697 struct cfs_rq *cfs_rq;
3698 struct sched_entity *se = &p->se;
3699 int task_sleep = flags & DEQUEUE_SLEEP;
3701 for_each_sched_entity(se) {
3702 cfs_rq = cfs_rq_of(se);
3703 dequeue_entity(cfs_rq, se, flags);
3706 * end evaluation on encountering a throttled cfs_rq
3708 * note: in the case of encountering a throttled cfs_rq we will
3709 * post the final h_nr_running decrement below.
3711 if (cfs_rq_throttled(cfs_rq))
3713 cfs_rq->h_nr_running--;
3715 /* Don't dequeue parent if it has other entities besides us */
3716 if (cfs_rq->load.weight) {
3718 * Bias pick_next to pick a task from this cfs_rq, as
3719 * p is sleeping when it is within its sched_slice.
3721 if (task_sleep && parent_entity(se))
3722 set_next_buddy(parent_entity(se));
3724 /* avoid re-evaluating load for this entity */
3725 se = parent_entity(se);
3728 flags |= DEQUEUE_SLEEP;
3731 for_each_sched_entity(se) {
3732 cfs_rq = cfs_rq_of(se);
3733 cfs_rq->h_nr_running--;
3735 if (cfs_rq_throttled(cfs_rq))
3738 update_cfs_shares(cfs_rq);
3739 update_entity_load_avg(se, 1);
3744 update_rq_runnable_avg(rq, 1);
3750 /* Used instead of source_load when we know the type == 0 */
3751 static unsigned long weighted_cpuload(const int cpu)
3753 return cpu_rq(cpu)->cfs.runnable_load_avg;
3757 * Return a low guess at the load of a migration-source cpu weighted
3758 * according to the scheduling class and "nice" value.
3760 * We want to under-estimate the load of migration sources, to
3761 * balance conservatively.
3763 static unsigned long source_load(int cpu, int type)
3765 struct rq *rq = cpu_rq(cpu);
3766 unsigned long total = weighted_cpuload(cpu);
3768 if (type == 0 || !sched_feat(LB_BIAS))
3771 return min(rq->cpu_load[type-1], total);
3775 * Return a high guess at the load of a migration-target cpu weighted
3776 * according to the scheduling class and "nice" value.
3778 static unsigned long target_load(int cpu, int type)
3780 struct rq *rq = cpu_rq(cpu);
3781 unsigned long total = weighted_cpuload(cpu);
3783 if (type == 0 || !sched_feat(LB_BIAS))
3786 return max(rq->cpu_load[type-1], total);
3789 static unsigned long power_of(int cpu)
3791 return cpu_rq(cpu)->cpu_power;
3794 static unsigned long cpu_avg_load_per_task(int cpu)
3796 struct rq *rq = cpu_rq(cpu);
3797 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3798 unsigned long load_avg = rq->cfs.runnable_load_avg;
3801 return load_avg / nr_running;
3806 static void record_wakee(struct task_struct *p)
3809 * Rough decay (wiping) for cost saving, don't worry
3810 * about the boundary, really active task won't care
3813 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3814 current->wakee_flips = 0;
3815 current->wakee_flip_decay_ts = jiffies;
3818 if (current->last_wakee != p) {
3819 current->last_wakee = p;
3820 current->wakee_flips++;
3824 static void task_waking_fair(struct task_struct *p)
3826 struct sched_entity *se = &p->se;
3827 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3830 #ifndef CONFIG_64BIT
3831 u64 min_vruntime_copy;
3834 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3836 min_vruntime = cfs_rq->min_vruntime;
3837 } while (min_vruntime != min_vruntime_copy);
3839 min_vruntime = cfs_rq->min_vruntime;
3842 se->vruntime -= min_vruntime;
3846 #ifdef CONFIG_FAIR_GROUP_SCHED
3848 * effective_load() calculates the load change as seen from the root_task_group
3850 * Adding load to a group doesn't make a group heavier, but can cause movement
3851 * of group shares between cpus. Assuming the shares were perfectly aligned one
3852 * can calculate the shift in shares.
3854 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3855 * on this @cpu and results in a total addition (subtraction) of @wg to the
3856 * total group weight.
3858 * Given a runqueue weight distribution (rw_i) we can compute a shares
3859 * distribution (s_i) using:
3861 * s_i = rw_i / \Sum rw_j (1)
3863 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3864 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3865 * shares distribution (s_i):
3867 * rw_i = { 2, 4, 1, 0 }
3868 * s_i = { 2/7, 4/7, 1/7, 0 }
3870 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3871 * task used to run on and the CPU the waker is running on), we need to
3872 * compute the effect of waking a task on either CPU and, in case of a sync
3873 * wakeup, compute the effect of the current task going to sleep.
3875 * So for a change of @wl to the local @cpu with an overall group weight change
3876 * of @wl we can compute the new shares distribution (s'_i) using:
3878 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3880 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3881 * differences in waking a task to CPU 0. The additional task changes the
3882 * weight and shares distributions like:
3884 * rw'_i = { 3, 4, 1, 0 }
3885 * s'_i = { 3/8, 4/8, 1/8, 0 }
3887 * We can then compute the difference in effective weight by using:
3889 * dw_i = S * (s'_i - s_i) (3)
3891 * Where 'S' is the group weight as seen by its parent.
3893 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3894 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3895 * 4/7) times the weight of the group.
3897 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3899 struct sched_entity *se = tg->se[cpu];
3901 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3904 for_each_sched_entity(se) {
3910 * W = @wg + \Sum rw_j
3912 W = wg + calc_tg_weight(tg, se->my_q);
3917 w = se->my_q->load.weight + wl;
3920 * wl = S * s'_i; see (2)
3923 wl = (w * tg->shares) / W;
3928 * Per the above, wl is the new se->load.weight value; since
3929 * those are clipped to [MIN_SHARES, ...) do so now. See
3930 * calc_cfs_shares().
3932 if (wl < MIN_SHARES)
3936 * wl = dw_i = S * (s'_i - s_i); see (3)
3938 wl -= se->load.weight;
3941 * Recursively apply this logic to all parent groups to compute
3942 * the final effective load change on the root group. Since
3943 * only the @tg group gets extra weight, all parent groups can
3944 * only redistribute existing shares. @wl is the shift in shares
3945 * resulting from this level per the above.
3954 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3961 static int wake_wide(struct task_struct *p)
3963 int factor = this_cpu_read(sd_llc_size);
3966 * Yeah, it's the switching-frequency, could means many wakee or
3967 * rapidly switch, use factor here will just help to automatically
3968 * adjust the loose-degree, so bigger node will lead to more pull.
3970 if (p->wakee_flips > factor) {
3972 * wakee is somewhat hot, it needs certain amount of cpu
3973 * resource, so if waker is far more hot, prefer to leave
3976 if (current->wakee_flips > (factor * p->wakee_flips))
3983 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3985 s64 this_load, load;
3986 int idx, this_cpu, prev_cpu;
3987 unsigned long tl_per_task;
3988 struct task_group *tg;
3989 unsigned long weight;
3993 * If we wake multiple tasks be careful to not bounce
3994 * ourselves around too much.
4000 this_cpu = smp_processor_id();
4001 prev_cpu = task_cpu(p);
4002 load = source_load(prev_cpu, idx);
4003 this_load = target_load(this_cpu, idx);
4006 * If sync wakeup then subtract the (maximum possible)
4007 * effect of the currently running task from the load
4008 * of the current CPU:
4011 tg = task_group(current);
4012 weight = current->se.load.weight;
4014 this_load += effective_load(tg, this_cpu, -weight, -weight);
4015 load += effective_load(tg, prev_cpu, 0, -weight);
4019 weight = p->se.load.weight;
4022 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4023 * due to the sync cause above having dropped this_load to 0, we'll
4024 * always have an imbalance, but there's really nothing you can do
4025 * about that, so that's good too.
4027 * Otherwise check if either cpus are near enough in load to allow this
4028 * task to be woken on this_cpu.
4030 if (this_load > 0) {
4031 s64 this_eff_load, prev_eff_load;
4033 this_eff_load = 100;
4034 this_eff_load *= power_of(prev_cpu);
4035 this_eff_load *= this_load +
4036 effective_load(tg, this_cpu, weight, weight);
4038 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4039 prev_eff_load *= power_of(this_cpu);
4040 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4042 balanced = this_eff_load <= prev_eff_load;
4047 * If the currently running task will sleep within
4048 * a reasonable amount of time then attract this newly
4051 if (sync && balanced)
4054 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4055 tl_per_task = cpu_avg_load_per_task(this_cpu);
4058 (this_load <= load &&
4059 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4061 * This domain has SD_WAKE_AFFINE and
4062 * p is cache cold in this domain, and
4063 * there is no bad imbalance.
4065 schedstat_inc(sd, ttwu_move_affine);
4066 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4074 * find_idlest_group finds and returns the least busy CPU group within the
4077 static struct sched_group *
4078 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4079 int this_cpu, int load_idx)
4081 struct sched_group *idlest = NULL, *group = sd->groups;
4082 unsigned long min_load = ULONG_MAX, this_load = 0;
4083 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4086 unsigned long load, avg_load;
4090 /* Skip over this group if it has no CPUs allowed */
4091 if (!cpumask_intersects(sched_group_cpus(group),
4092 tsk_cpus_allowed(p)))
4095 local_group = cpumask_test_cpu(this_cpu,
4096 sched_group_cpus(group));
4098 /* Tally up the load of all CPUs in the group */
4101 for_each_cpu(i, sched_group_cpus(group)) {
4102 /* Bias balancing toward cpus of our domain */
4104 load = source_load(i, load_idx);
4106 load = target_load(i, load_idx);
4111 /* Adjust by relative CPU power of the group */
4112 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4115 this_load = avg_load;
4116 } else if (avg_load < min_load) {
4117 min_load = avg_load;
4120 } while (group = group->next, group != sd->groups);
4122 if (!idlest || 100*this_load < imbalance*min_load)
4128 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4131 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4133 unsigned long load, min_load = ULONG_MAX;
4137 /* Traverse only the allowed CPUs */
4138 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4139 load = weighted_cpuload(i);
4141 if (load < min_load || (load == min_load && i == this_cpu)) {
4151 * Try and locate an idle CPU in the sched_domain.
4153 static int select_idle_sibling(struct task_struct *p, int target)
4155 struct sched_domain *sd;
4156 struct sched_group *sg;
4157 int i = task_cpu(p);
4159 if (idle_cpu(target))
4163 * If the prevous cpu is cache affine and idle, don't be stupid.
4165 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4169 * Otherwise, iterate the domains and find an elegible idle cpu.
4171 sd = rcu_dereference(per_cpu(sd_llc, target));
4172 for_each_lower_domain(sd) {
4175 if (!cpumask_intersects(sched_group_cpus(sg),
4176 tsk_cpus_allowed(p)))
4179 for_each_cpu(i, sched_group_cpus(sg)) {
4180 if (i == target || !idle_cpu(i))
4184 target = cpumask_first_and(sched_group_cpus(sg),
4185 tsk_cpus_allowed(p));
4189 } while (sg != sd->groups);
4196 * sched_balance_self: balance the current task (running on cpu) in domains
4197 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4200 * Balance, ie. select the least loaded group.
4202 * Returns the target CPU number, or the same CPU if no balancing is needed.
4204 * preempt must be disabled.
4207 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4209 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4210 int cpu = smp_processor_id();
4212 int want_affine = 0;
4213 int sync = wake_flags & WF_SYNC;
4215 if (p->nr_cpus_allowed == 1)
4218 if (sd_flag & SD_BALANCE_WAKE) {
4219 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4225 for_each_domain(cpu, tmp) {
4226 if (!(tmp->flags & SD_LOAD_BALANCE))
4230 * If both cpu and prev_cpu are part of this domain,
4231 * cpu is a valid SD_WAKE_AFFINE target.
4233 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4234 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4239 if (tmp->flags & sd_flag)
4244 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4247 new_cpu = select_idle_sibling(p, prev_cpu);
4252 int load_idx = sd->forkexec_idx;
4253 struct sched_group *group;
4256 if (!(sd->flags & sd_flag)) {
4261 if (sd_flag & SD_BALANCE_WAKE)
4262 load_idx = sd->wake_idx;
4264 group = find_idlest_group(sd, p, cpu, load_idx);
4270 new_cpu = find_idlest_cpu(group, p, cpu);
4271 if (new_cpu == -1 || new_cpu == cpu) {
4272 /* Now try balancing at a lower domain level of cpu */
4277 /* Now try balancing at a lower domain level of new_cpu */
4279 weight = sd->span_weight;
4281 for_each_domain(cpu, tmp) {
4282 if (weight <= tmp->span_weight)
4284 if (tmp->flags & sd_flag)
4287 /* while loop will break here if sd == NULL */
4296 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4297 * cfs_rq_of(p) references at time of call are still valid and identify the
4298 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4299 * other assumptions, including the state of rq->lock, should be made.
4302 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4304 struct sched_entity *se = &p->se;
4305 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4308 * Load tracking: accumulate removed load so that it can be processed
4309 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4310 * to blocked load iff they have a positive decay-count. It can never
4311 * be negative here since on-rq tasks have decay-count == 0.
4313 if (se->avg.decay_count) {
4314 se->avg.decay_count = -__synchronize_entity_decay(se);
4315 atomic_long_add(se->avg.load_avg_contrib,
4316 &cfs_rq->removed_load);
4319 #endif /* CONFIG_SMP */
4321 static unsigned long
4322 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4324 unsigned long gran = sysctl_sched_wakeup_granularity;
4327 * Since its curr running now, convert the gran from real-time
4328 * to virtual-time in his units.
4330 * By using 'se' instead of 'curr' we penalize light tasks, so
4331 * they get preempted easier. That is, if 'se' < 'curr' then
4332 * the resulting gran will be larger, therefore penalizing the
4333 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4334 * be smaller, again penalizing the lighter task.
4336 * This is especially important for buddies when the leftmost
4337 * task is higher priority than the buddy.
4339 return calc_delta_fair(gran, se);
4343 * Should 'se' preempt 'curr'.
4357 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4359 s64 gran, vdiff = curr->vruntime - se->vruntime;
4364 gran = wakeup_gran(curr, se);
4371 static void set_last_buddy(struct sched_entity *se)
4373 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4376 for_each_sched_entity(se)
4377 cfs_rq_of(se)->last = se;
4380 static void set_next_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)->next = se;
4389 static void set_skip_buddy(struct sched_entity *se)
4391 for_each_sched_entity(se)
4392 cfs_rq_of(se)->skip = se;
4396 * Preempt the current task with a newly woken task if needed:
4398 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4400 struct task_struct *curr = rq->curr;
4401 struct sched_entity *se = &curr->se, *pse = &p->se;
4402 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4403 int scale = cfs_rq->nr_running >= sched_nr_latency;
4404 int next_buddy_marked = 0;
4406 if (unlikely(se == pse))
4410 * This is possible from callers such as move_task(), in which we
4411 * unconditionally check_prempt_curr() after an enqueue (which may have
4412 * lead to a throttle). This both saves work and prevents false
4413 * next-buddy nomination below.
4415 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4418 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4419 set_next_buddy(pse);
4420 next_buddy_marked = 1;
4424 * We can come here with TIF_NEED_RESCHED already set from new task
4427 * Note: this also catches the edge-case of curr being in a throttled
4428 * group (e.g. via set_curr_task), since update_curr() (in the
4429 * enqueue of curr) will have resulted in resched being set. This
4430 * prevents us from potentially nominating it as a false LAST_BUDDY
4433 if (test_tsk_need_resched(curr))
4436 /* Idle tasks are by definition preempted by non-idle tasks. */
4437 if (unlikely(curr->policy == SCHED_IDLE) &&
4438 likely(p->policy != SCHED_IDLE))
4442 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4443 * is driven by the tick):
4445 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4448 find_matching_se(&se, &pse);
4449 update_curr(cfs_rq_of(se));
4451 if (wakeup_preempt_entity(se, pse) == 1) {
4453 * Bias pick_next to pick the sched entity that is
4454 * triggering this preemption.
4456 if (!next_buddy_marked)
4457 set_next_buddy(pse);
4466 * Only set the backward buddy when the current task is still
4467 * on the rq. This can happen when a wakeup gets interleaved
4468 * with schedule on the ->pre_schedule() or idle_balance()
4469 * point, either of which can * drop the rq lock.
4471 * Also, during early boot the idle thread is in the fair class,
4472 * for obvious reasons its a bad idea to schedule back to it.
4474 if (unlikely(!se->on_rq || curr == rq->idle))
4477 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4481 static struct task_struct *pick_next_task_fair(struct rq *rq)
4483 struct task_struct *p;
4484 struct cfs_rq *cfs_rq = &rq->cfs;
4485 struct sched_entity *se;
4487 if (!cfs_rq->nr_running)
4491 se = pick_next_entity(cfs_rq);
4492 set_next_entity(cfs_rq, se);
4493 cfs_rq = group_cfs_rq(se);
4497 if (hrtick_enabled(rq))
4498 hrtick_start_fair(rq, p);
4504 * Account for a descheduled task:
4506 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4508 struct sched_entity *se = &prev->se;
4509 struct cfs_rq *cfs_rq;
4511 for_each_sched_entity(se) {
4512 cfs_rq = cfs_rq_of(se);
4513 put_prev_entity(cfs_rq, se);
4518 * sched_yield() is very simple
4520 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4522 static void yield_task_fair(struct rq *rq)
4524 struct task_struct *curr = rq->curr;
4525 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4526 struct sched_entity *se = &curr->se;
4529 * Are we the only task in the tree?
4531 if (unlikely(rq->nr_running == 1))
4534 clear_buddies(cfs_rq, se);
4536 if (curr->policy != SCHED_BATCH) {
4537 update_rq_clock(rq);
4539 * Update run-time statistics of the 'current'.
4541 update_curr(cfs_rq);
4543 * Tell update_rq_clock() that we've just updated,
4544 * so we don't do microscopic update in schedule()
4545 * and double the fastpath cost.
4547 rq->skip_clock_update = 1;
4553 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4555 struct sched_entity *se = &p->se;
4557 /* throttled hierarchies are not runnable */
4558 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4561 /* Tell the scheduler that we'd really like pse to run next. */
4564 yield_task_fair(rq);
4570 /**************************************************
4571 * Fair scheduling class load-balancing methods.
4575 * The purpose of load-balancing is to achieve the same basic fairness the
4576 * per-cpu scheduler provides, namely provide a proportional amount of compute
4577 * time to each task. This is expressed in the following equation:
4579 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4581 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4582 * W_i,0 is defined as:
4584 * W_i,0 = \Sum_j w_i,j (2)
4586 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4587 * is derived from the nice value as per prio_to_weight[].
4589 * The weight average is an exponential decay average of the instantaneous
4592 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4594 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4595 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4596 * can also include other factors [XXX].
4598 * To achieve this balance we define a measure of imbalance which follows
4599 * directly from (1):
4601 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4603 * We them move tasks around to minimize the imbalance. In the continuous
4604 * function space it is obvious this converges, in the discrete case we get
4605 * a few fun cases generally called infeasible weight scenarios.
4608 * - infeasible weights;
4609 * - local vs global optima in the discrete case. ]
4614 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4615 * for all i,j solution, we create a tree of cpus that follows the hardware
4616 * topology where each level pairs two lower groups (or better). This results
4617 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4618 * tree to only the first of the previous level and we decrease the frequency
4619 * of load-balance at each level inv. proportional to the number of cpus in
4625 * \Sum { --- * --- * 2^i } = O(n) (5)
4627 * `- size of each group
4628 * | | `- number of cpus doing load-balance
4630 * `- sum over all levels
4632 * Coupled with a limit on how many tasks we can migrate every balance pass,
4633 * this makes (5) the runtime complexity of the balancer.
4635 * An important property here is that each CPU is still (indirectly) connected
4636 * to every other cpu in at most O(log n) steps:
4638 * The adjacency matrix of the resulting graph is given by:
4641 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4644 * And you'll find that:
4646 * A^(log_2 n)_i,j != 0 for all i,j (7)
4648 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4649 * The task movement gives a factor of O(m), giving a convergence complexity
4652 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4657 * In order to avoid CPUs going idle while there's still work to do, new idle
4658 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4659 * tree itself instead of relying on other CPUs to bring it work.
4661 * This adds some complexity to both (5) and (8) but it reduces the total idle
4669 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4672 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4677 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4679 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4681 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4684 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4685 * rewrite all of this once again.]
4688 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4690 enum fbq_type { regular, remote, all };
4692 #define LBF_ALL_PINNED 0x01
4693 #define LBF_NEED_BREAK 0x02
4694 #define LBF_DST_PINNED 0x04
4695 #define LBF_SOME_PINNED 0x08
4698 struct sched_domain *sd;
4706 struct cpumask *dst_grpmask;
4708 enum cpu_idle_type idle;
4710 /* The set of CPUs under consideration for load-balancing */
4711 struct cpumask *cpus;
4716 unsigned int loop_break;
4717 unsigned int loop_max;
4719 enum fbq_type fbq_type;
4723 * move_task - move a task from one runqueue to another runqueue.
4724 * Both runqueues must be locked.
4726 static void move_task(struct task_struct *p, struct lb_env *env)
4728 deactivate_task(env->src_rq, p, 0);
4729 set_task_cpu(p, env->dst_cpu);
4730 activate_task(env->dst_rq, p, 0);
4731 check_preempt_curr(env->dst_rq, p, 0);
4735 * Is this task likely cache-hot:
4738 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4742 if (p->sched_class != &fair_sched_class)
4745 if (unlikely(p->policy == SCHED_IDLE))
4749 * Buddy candidates are cache hot:
4751 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4752 (&p->se == cfs_rq_of(&p->se)->next ||
4753 &p->se == cfs_rq_of(&p->se)->last))
4756 if (sysctl_sched_migration_cost == -1)
4758 if (sysctl_sched_migration_cost == 0)
4761 delta = now - p->se.exec_start;
4763 return delta < (s64)sysctl_sched_migration_cost;
4766 #ifdef CONFIG_NUMA_BALANCING
4767 /* Returns true if the destination node has incurred more faults */
4768 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4770 int src_nid, dst_nid;
4772 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4773 !(env->sd->flags & SD_NUMA)) {
4777 src_nid = cpu_to_node(env->src_cpu);
4778 dst_nid = cpu_to_node(env->dst_cpu);
4780 if (src_nid == dst_nid)
4783 /* Always encourage migration to the preferred node. */
4784 if (dst_nid == p->numa_preferred_nid)
4787 /* If both task and group weight improve, this move is a winner. */
4788 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4789 group_weight(p, dst_nid) > group_weight(p, src_nid))
4796 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4798 int src_nid, dst_nid;
4800 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4803 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4806 src_nid = cpu_to_node(env->src_cpu);
4807 dst_nid = cpu_to_node(env->dst_cpu);
4809 if (src_nid == dst_nid)
4812 /* Migrating away from the preferred node is always bad. */
4813 if (src_nid == p->numa_preferred_nid)
4816 /* If either task or group weight get worse, don't do it. */
4817 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4818 group_weight(p, dst_nid) < group_weight(p, src_nid))
4825 static inline bool migrate_improves_locality(struct task_struct *p,
4831 static inline bool migrate_degrades_locality(struct task_struct *p,
4839 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4842 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4844 int tsk_cache_hot = 0;
4846 * We do not migrate tasks that are:
4847 * 1) throttled_lb_pair, or
4848 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4849 * 3) running (obviously), or
4850 * 4) are cache-hot on their current CPU.
4852 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4855 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4858 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4860 env->flags |= LBF_SOME_PINNED;
4863 * Remember if this task can be migrated to any other cpu in
4864 * our sched_group. We may want to revisit it if we couldn't
4865 * meet load balance goals by pulling other tasks on src_cpu.
4867 * Also avoid computing new_dst_cpu if we have already computed
4868 * one in current iteration.
4870 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4873 /* Prevent to re-select dst_cpu via env's cpus */
4874 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4875 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4876 env->flags |= LBF_DST_PINNED;
4877 env->new_dst_cpu = cpu;
4885 /* Record that we found atleast one task that could run on dst_cpu */
4886 env->flags &= ~LBF_ALL_PINNED;
4888 if (task_running(env->src_rq, p)) {
4889 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4894 * Aggressive migration if:
4895 * 1) destination numa is preferred
4896 * 2) task is cache cold, or
4897 * 3) too many balance attempts have failed.
4899 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4901 tsk_cache_hot = migrate_degrades_locality(p, env);
4903 if (migrate_improves_locality(p, env)) {
4904 #ifdef CONFIG_SCHEDSTATS
4905 if (tsk_cache_hot) {
4906 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4907 schedstat_inc(p, se.statistics.nr_forced_migrations);
4913 if (!tsk_cache_hot ||
4914 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4916 if (tsk_cache_hot) {
4917 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4918 schedstat_inc(p, se.statistics.nr_forced_migrations);
4924 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4929 * move_one_task tries to move exactly one task from busiest to this_rq, as
4930 * part of active balancing operations within "domain".
4931 * Returns 1 if successful and 0 otherwise.
4933 * Called with both runqueues locked.
4935 static int move_one_task(struct lb_env *env)
4937 struct task_struct *p, *n;
4939 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4940 if (!can_migrate_task(p, env))
4945 * Right now, this is only the second place move_task()
4946 * is called, so we can safely collect move_task()
4947 * stats here rather than inside move_task().
4949 schedstat_inc(env->sd, lb_gained[env->idle]);
4955 static const unsigned int sched_nr_migrate_break = 32;
4958 * move_tasks tries to move up to imbalance weighted load from busiest to
4959 * this_rq, as part of a balancing operation within domain "sd".
4960 * Returns 1 if successful and 0 otherwise.
4962 * Called with both runqueues locked.
4964 static int move_tasks(struct lb_env *env)
4966 struct list_head *tasks = &env->src_rq->cfs_tasks;
4967 struct task_struct *p;
4971 if (env->imbalance <= 0)
4974 while (!list_empty(tasks)) {
4975 p = list_first_entry(tasks, struct task_struct, se.group_node);
4978 /* We've more or less seen every task there is, call it quits */
4979 if (env->loop > env->loop_max)
4982 /* take a breather every nr_migrate tasks */
4983 if (env->loop > env->loop_break) {
4984 env->loop_break += sched_nr_migrate_break;
4985 env->flags |= LBF_NEED_BREAK;
4989 if (!can_migrate_task(p, env))
4992 load = task_h_load(p);
4994 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4997 if ((load / 2) > env->imbalance)
5002 env->imbalance -= load;
5004 #ifdef CONFIG_PREEMPT
5006 * NEWIDLE balancing is a source of latency, so preemptible
5007 * kernels will stop after the first task is pulled to minimize
5008 * the critical section.
5010 if (env->idle == CPU_NEWLY_IDLE)
5015 * We only want to steal up to the prescribed amount of
5018 if (env->imbalance <= 0)
5023 list_move_tail(&p->se.group_node, tasks);
5027 * Right now, this is one of only two places move_task() is called,
5028 * so we can safely collect move_task() stats here rather than
5029 * inside move_task().
5031 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5036 #ifdef CONFIG_FAIR_GROUP_SCHED
5038 * update tg->load_weight by folding this cpu's load_avg
5040 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5042 struct sched_entity *se = tg->se[cpu];
5043 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5045 /* throttled entities do not contribute to load */
5046 if (throttled_hierarchy(cfs_rq))
5049 update_cfs_rq_blocked_load(cfs_rq, 1);
5052 update_entity_load_avg(se, 1);
5054 * We pivot on our runnable average having decayed to zero for
5055 * list removal. This generally implies that all our children
5056 * have also been removed (modulo rounding error or bandwidth
5057 * control); however, such cases are rare and we can fix these
5060 * TODO: fix up out-of-order children on enqueue.
5062 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5063 list_del_leaf_cfs_rq(cfs_rq);
5065 struct rq *rq = rq_of(cfs_rq);
5066 update_rq_runnable_avg(rq, rq->nr_running);
5070 static void update_blocked_averages(int cpu)
5072 struct rq *rq = cpu_rq(cpu);
5073 struct cfs_rq *cfs_rq;
5074 unsigned long flags;
5076 raw_spin_lock_irqsave(&rq->lock, flags);
5077 update_rq_clock(rq);
5079 * Iterates the task_group tree in a bottom up fashion, see
5080 * list_add_leaf_cfs_rq() for details.
5082 for_each_leaf_cfs_rq(rq, cfs_rq) {
5084 * Note: We may want to consider periodically releasing
5085 * rq->lock about these updates so that creating many task
5086 * groups does not result in continually extending hold time.
5088 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5091 raw_spin_unlock_irqrestore(&rq->lock, flags);
5095 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5096 * This needs to be done in a top-down fashion because the load of a child
5097 * group is a fraction of its parents load.
5099 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5101 struct rq *rq = rq_of(cfs_rq);
5102 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5103 unsigned long now = jiffies;
5106 if (cfs_rq->last_h_load_update == now)
5109 cfs_rq->h_load_next = NULL;
5110 for_each_sched_entity(se) {
5111 cfs_rq = cfs_rq_of(se);
5112 cfs_rq->h_load_next = se;
5113 if (cfs_rq->last_h_load_update == now)
5118 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5119 cfs_rq->last_h_load_update = now;
5122 while ((se = cfs_rq->h_load_next) != NULL) {
5123 load = cfs_rq->h_load;
5124 load = div64_ul(load * se->avg.load_avg_contrib,
5125 cfs_rq->runnable_load_avg + 1);
5126 cfs_rq = group_cfs_rq(se);
5127 cfs_rq->h_load = load;
5128 cfs_rq->last_h_load_update = now;
5132 static unsigned long task_h_load(struct task_struct *p)
5134 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5136 update_cfs_rq_h_load(cfs_rq);
5137 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5138 cfs_rq->runnable_load_avg + 1);
5141 static inline void update_blocked_averages(int cpu)
5145 static unsigned long task_h_load(struct task_struct *p)
5147 return p->se.avg.load_avg_contrib;
5151 /********** Helpers for find_busiest_group ************************/
5153 * sg_lb_stats - stats of a sched_group required for load_balancing
5155 struct sg_lb_stats {
5156 unsigned long avg_load; /*Avg load across the CPUs of the group */
5157 unsigned long group_load; /* Total load over the CPUs of the group */
5158 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5159 unsigned long load_per_task;
5160 unsigned long group_power;
5161 unsigned int sum_nr_running; /* Nr tasks running in the group */
5162 unsigned int group_capacity;
5163 unsigned int idle_cpus;
5164 unsigned int group_weight;
5165 int group_imb; /* Is there an imbalance in the group ? */
5166 int group_has_capacity; /* Is there extra capacity in the group? */
5167 #ifdef CONFIG_NUMA_BALANCING
5168 unsigned int nr_numa_running;
5169 unsigned int nr_preferred_running;
5174 * sd_lb_stats - Structure to store the statistics of a sched_domain
5175 * during load balancing.
5177 struct sd_lb_stats {
5178 struct sched_group *busiest; /* Busiest group in this sd */
5179 struct sched_group *local; /* Local group in this sd */
5180 unsigned long total_load; /* Total load of all groups in sd */
5181 unsigned long total_pwr; /* Total power of all groups in sd */
5182 unsigned long avg_load; /* Average load across all groups in sd */
5184 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5185 struct sg_lb_stats local_stat; /* Statistics of the local group */
5188 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5191 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5192 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5193 * We must however clear busiest_stat::avg_load because
5194 * update_sd_pick_busiest() reads this before assignment.
5196 *sds = (struct sd_lb_stats){
5208 * get_sd_load_idx - Obtain the load index for a given sched domain.
5209 * @sd: The sched_domain whose load_idx is to be obtained.
5210 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5212 * Return: The load index.
5214 static inline int get_sd_load_idx(struct sched_domain *sd,
5215 enum cpu_idle_type idle)
5221 load_idx = sd->busy_idx;
5224 case CPU_NEWLY_IDLE:
5225 load_idx = sd->newidle_idx;
5228 load_idx = sd->idle_idx;
5235 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5237 return SCHED_POWER_SCALE;
5240 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5242 return default_scale_freq_power(sd, cpu);
5245 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5247 unsigned long weight = sd->span_weight;
5248 unsigned long smt_gain = sd->smt_gain;
5255 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5257 return default_scale_smt_power(sd, cpu);
5260 static unsigned long scale_rt_power(int cpu)
5262 struct rq *rq = cpu_rq(cpu);
5263 u64 total, available, age_stamp, avg;
5266 * Since we're reading these variables without serialization make sure
5267 * we read them once before doing sanity checks on them.
5269 age_stamp = ACCESS_ONCE(rq->age_stamp);
5270 avg = ACCESS_ONCE(rq->rt_avg);
5272 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5274 if (unlikely(total < avg)) {
5275 /* Ensures that power won't end up being negative */
5278 available = total - avg;
5281 if (unlikely((s64)total < SCHED_POWER_SCALE))
5282 total = SCHED_POWER_SCALE;
5284 total >>= SCHED_POWER_SHIFT;
5286 return div_u64(available, total);
5289 static void update_cpu_power(struct sched_domain *sd, int cpu)
5291 unsigned long weight = sd->span_weight;
5292 unsigned long power = SCHED_POWER_SCALE;
5293 struct sched_group *sdg = sd->groups;
5295 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5296 if (sched_feat(ARCH_POWER))
5297 power *= arch_scale_smt_power(sd, cpu);
5299 power *= default_scale_smt_power(sd, cpu);
5301 power >>= SCHED_POWER_SHIFT;
5304 sdg->sgp->power_orig = power;
5306 if (sched_feat(ARCH_POWER))
5307 power *= arch_scale_freq_power(sd, cpu);
5309 power *= default_scale_freq_power(sd, cpu);
5311 power >>= SCHED_POWER_SHIFT;
5313 power *= scale_rt_power(cpu);
5314 power >>= SCHED_POWER_SHIFT;
5319 cpu_rq(cpu)->cpu_power = power;
5320 sdg->sgp->power = power;
5323 void update_group_power(struct sched_domain *sd, int cpu)
5325 struct sched_domain *child = sd->child;
5326 struct sched_group *group, *sdg = sd->groups;
5327 unsigned long power, power_orig;
5328 unsigned long interval;
5330 interval = msecs_to_jiffies(sd->balance_interval);
5331 interval = clamp(interval, 1UL, max_load_balance_interval);
5332 sdg->sgp->next_update = jiffies + interval;
5335 update_cpu_power(sd, cpu);
5339 power_orig = power = 0;
5341 if (child->flags & SD_OVERLAP) {
5343 * SD_OVERLAP domains cannot assume that child groups
5344 * span the current group.
5347 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5348 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5350 power_orig += sg->sgp->power_orig;
5351 power += sg->sgp->power;
5355 * !SD_OVERLAP domains can assume that child groups
5356 * span the current group.
5359 group = child->groups;
5361 power_orig += group->sgp->power_orig;
5362 power += group->sgp->power;
5363 group = group->next;
5364 } while (group != child->groups);
5367 sdg->sgp->power_orig = power_orig;
5368 sdg->sgp->power = power;
5372 * Try and fix up capacity for tiny siblings, this is needed when
5373 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5374 * which on its own isn't powerful enough.
5376 * See update_sd_pick_busiest() and check_asym_packing().
5379 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5382 * Only siblings can have significantly less than SCHED_POWER_SCALE
5384 if (!(sd->flags & SD_SHARE_CPUPOWER))
5388 * If ~90% of the cpu_power is still there, we're good.
5390 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5397 * Group imbalance indicates (and tries to solve) the problem where balancing
5398 * groups is inadequate due to tsk_cpus_allowed() constraints.
5400 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5401 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5404 * { 0 1 2 3 } { 4 5 6 7 }
5407 * If we were to balance group-wise we'd place two tasks in the first group and
5408 * two tasks in the second group. Clearly this is undesired as it will overload
5409 * cpu 3 and leave one of the cpus in the second group unused.
5411 * The current solution to this issue is detecting the skew in the first group
5412 * by noticing the lower domain failed to reach balance and had difficulty
5413 * moving tasks due to affinity constraints.
5415 * When this is so detected; this group becomes a candidate for busiest; see
5416 * update_sd_pick_busiest(). And calculcate_imbalance() and
5417 * find_busiest_group() avoid some of the usual balance conditions to allow it
5418 * to create an effective group imbalance.
5420 * This is a somewhat tricky proposition since the next run might not find the
5421 * group imbalance and decide the groups need to be balanced again. A most
5422 * subtle and fragile situation.
5425 static inline int sg_imbalanced(struct sched_group *group)
5427 return group->sgp->imbalance;
5431 * Compute the group capacity.
5433 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5434 * first dividing out the smt factor and computing the actual number of cores
5435 * and limit power unit capacity with that.
5437 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5439 unsigned int capacity, smt, cpus;
5440 unsigned int power, power_orig;
5442 power = group->sgp->power;
5443 power_orig = group->sgp->power_orig;
5444 cpus = group->group_weight;
5446 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5447 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5448 capacity = cpus / smt; /* cores */
5450 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5452 capacity = fix_small_capacity(env->sd, group);
5458 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5459 * @env: The load balancing environment.
5460 * @group: sched_group whose statistics are to be updated.
5461 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5462 * @local_group: Does group contain this_cpu.
5463 * @sgs: variable to hold the statistics for this group.
5465 static inline void update_sg_lb_stats(struct lb_env *env,
5466 struct sched_group *group, int load_idx,
5467 int local_group, struct sg_lb_stats *sgs)
5469 unsigned long nr_running;
5473 memset(sgs, 0, sizeof(*sgs));
5475 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5476 struct rq *rq = cpu_rq(i);
5478 nr_running = rq->nr_running;
5480 /* Bias balancing toward cpus of our domain */
5482 load = target_load(i, load_idx);
5484 load = source_load(i, load_idx);
5486 sgs->group_load += load;
5487 sgs->sum_nr_running += nr_running;
5488 #ifdef CONFIG_NUMA_BALANCING
5489 sgs->nr_numa_running += rq->nr_numa_running;
5490 sgs->nr_preferred_running += rq->nr_preferred_running;
5492 sgs->sum_weighted_load += weighted_cpuload(i);
5497 /* Adjust by relative CPU power of the group */
5498 sgs->group_power = group->sgp->power;
5499 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5501 if (sgs->sum_nr_running)
5502 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5504 sgs->group_weight = group->group_weight;
5506 sgs->group_imb = sg_imbalanced(group);
5507 sgs->group_capacity = sg_capacity(env, group);
5509 if (sgs->group_capacity > sgs->sum_nr_running)
5510 sgs->group_has_capacity = 1;
5514 * update_sd_pick_busiest - return 1 on busiest group
5515 * @env: The load balancing environment.
5516 * @sds: sched_domain statistics
5517 * @sg: sched_group candidate to be checked for being the busiest
5518 * @sgs: sched_group statistics
5520 * Determine if @sg is a busier group than the previously selected
5523 * Return: %true if @sg is a busier group than the previously selected
5524 * busiest group. %false otherwise.
5526 static bool update_sd_pick_busiest(struct lb_env *env,
5527 struct sd_lb_stats *sds,
5528 struct sched_group *sg,
5529 struct sg_lb_stats *sgs)
5531 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5534 if (sgs->sum_nr_running > sgs->group_capacity)
5541 * ASYM_PACKING needs to move all the work to the lowest
5542 * numbered CPUs in the group, therefore mark all groups
5543 * higher than ourself as busy.
5545 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5546 env->dst_cpu < group_first_cpu(sg)) {
5550 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5557 #ifdef CONFIG_NUMA_BALANCING
5558 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5560 if (sgs->sum_nr_running > sgs->nr_numa_running)
5562 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5567 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5569 if (rq->nr_running > rq->nr_numa_running)
5571 if (rq->nr_running > rq->nr_preferred_running)
5576 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5581 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5585 #endif /* CONFIG_NUMA_BALANCING */
5588 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5589 * @env: The load balancing environment.
5590 * @balance: Should we balance.
5591 * @sds: variable to hold the statistics for this sched_domain.
5593 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5595 struct sched_domain *child = env->sd->child;
5596 struct sched_group *sg = env->sd->groups;
5597 struct sg_lb_stats tmp_sgs;
5598 int load_idx, prefer_sibling = 0;
5600 if (child && child->flags & SD_PREFER_SIBLING)
5603 load_idx = get_sd_load_idx(env->sd, env->idle);
5606 struct sg_lb_stats *sgs = &tmp_sgs;
5609 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5612 sgs = &sds->local_stat;
5614 if (env->idle != CPU_NEWLY_IDLE ||
5615 time_after_eq(jiffies, sg->sgp->next_update))
5616 update_group_power(env->sd, env->dst_cpu);
5619 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5625 * In case the child domain prefers tasks go to siblings
5626 * first, lower the sg capacity to one so that we'll try
5627 * and move all the excess tasks away. We lower the capacity
5628 * of a group only if the local group has the capacity to fit
5629 * these excess tasks, i.e. nr_running < group_capacity. The
5630 * extra check prevents the case where you always pull from the
5631 * heaviest group when it is already under-utilized (possible
5632 * with a large weight task outweighs the tasks on the system).
5634 if (prefer_sibling && sds->local &&
5635 sds->local_stat.group_has_capacity)
5636 sgs->group_capacity = min(sgs->group_capacity, 1U);
5638 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5640 sds->busiest_stat = *sgs;
5644 /* Now, start updating sd_lb_stats */
5645 sds->total_load += sgs->group_load;
5646 sds->total_pwr += sgs->group_power;
5649 } while (sg != env->sd->groups);
5651 if (env->sd->flags & SD_NUMA)
5652 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5656 * check_asym_packing - Check to see if the group is packed into the
5659 * This is primarily intended to used at the sibling level. Some
5660 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5661 * case of POWER7, it can move to lower SMT modes only when higher
5662 * threads are idle. When in lower SMT modes, the threads will
5663 * perform better since they share less core resources. Hence when we
5664 * have idle threads, we want them to be the higher ones.
5666 * This packing function is run on idle threads. It checks to see if
5667 * the busiest CPU in this domain (core in the P7 case) has a higher
5668 * CPU number than the packing function is being run on. Here we are
5669 * assuming lower CPU number will be equivalent to lower a SMT thread
5672 * Return: 1 when packing is required and a task should be moved to
5673 * this CPU. The amount of the imbalance is returned in *imbalance.
5675 * @env: The load balancing environment.
5676 * @sds: Statistics of the sched_domain which is to be packed
5678 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5682 if (!(env->sd->flags & SD_ASYM_PACKING))
5688 busiest_cpu = group_first_cpu(sds->busiest);
5689 if (env->dst_cpu > busiest_cpu)
5692 env->imbalance = DIV_ROUND_CLOSEST(
5693 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5700 * fix_small_imbalance - Calculate the minor imbalance that exists
5701 * amongst the groups of a sched_domain, during
5703 * @env: The load balancing environment.
5704 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5707 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5709 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5710 unsigned int imbn = 2;
5711 unsigned long scaled_busy_load_per_task;
5712 struct sg_lb_stats *local, *busiest;
5714 local = &sds->local_stat;
5715 busiest = &sds->busiest_stat;
5717 if (!local->sum_nr_running)
5718 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5719 else if (busiest->load_per_task > local->load_per_task)
5722 scaled_busy_load_per_task =
5723 (busiest->load_per_task * SCHED_POWER_SCALE) /
5724 busiest->group_power;
5726 if (busiest->avg_load + scaled_busy_load_per_task >=
5727 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5728 env->imbalance = busiest->load_per_task;
5733 * OK, we don't have enough imbalance to justify moving tasks,
5734 * however we may be able to increase total CPU power used by
5738 pwr_now += busiest->group_power *
5739 min(busiest->load_per_task, busiest->avg_load);
5740 pwr_now += local->group_power *
5741 min(local->load_per_task, local->avg_load);
5742 pwr_now /= SCHED_POWER_SCALE;
5744 /* Amount of load we'd subtract */
5745 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5746 busiest->group_power;
5747 if (busiest->avg_load > tmp) {
5748 pwr_move += busiest->group_power *
5749 min(busiest->load_per_task,
5750 busiest->avg_load - tmp);
5753 /* Amount of load we'd add */
5754 if (busiest->avg_load * busiest->group_power <
5755 busiest->load_per_task * SCHED_POWER_SCALE) {
5756 tmp = (busiest->avg_load * busiest->group_power) /
5759 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5762 pwr_move += local->group_power *
5763 min(local->load_per_task, local->avg_load + tmp);
5764 pwr_move /= SCHED_POWER_SCALE;
5766 /* Move if we gain throughput */
5767 if (pwr_move > pwr_now)
5768 env->imbalance = busiest->load_per_task;
5772 * calculate_imbalance - Calculate the amount of imbalance present within the
5773 * groups of a given sched_domain during load balance.
5774 * @env: load balance environment
5775 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5777 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5779 unsigned long max_pull, load_above_capacity = ~0UL;
5780 struct sg_lb_stats *local, *busiest;
5782 local = &sds->local_stat;
5783 busiest = &sds->busiest_stat;
5785 if (busiest->group_imb) {
5787 * In the group_imb case we cannot rely on group-wide averages
5788 * to ensure cpu-load equilibrium, look at wider averages. XXX
5790 busiest->load_per_task =
5791 min(busiest->load_per_task, sds->avg_load);
5795 * In the presence of smp nice balancing, certain scenarios can have
5796 * max load less than avg load(as we skip the groups at or below
5797 * its cpu_power, while calculating max_load..)
5799 if (busiest->avg_load <= sds->avg_load ||
5800 local->avg_load >= sds->avg_load) {
5802 return fix_small_imbalance(env, sds);
5805 if (!busiest->group_imb) {
5807 * Don't want to pull so many tasks that a group would go idle.
5808 * Except of course for the group_imb case, since then we might
5809 * have to drop below capacity to reach cpu-load equilibrium.
5811 load_above_capacity =
5812 (busiest->sum_nr_running - busiest->group_capacity);
5814 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5815 load_above_capacity /= busiest->group_power;
5819 * We're trying to get all the cpus to the average_load, so we don't
5820 * want to push ourselves above the average load, nor do we wish to
5821 * reduce the max loaded cpu below the average load. At the same time,
5822 * we also don't want to reduce the group load below the group capacity
5823 * (so that we can implement power-savings policies etc). Thus we look
5824 * for the minimum possible imbalance.
5826 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5828 /* How much load to actually move to equalise the imbalance */
5829 env->imbalance = min(
5830 max_pull * busiest->group_power,
5831 (sds->avg_load - local->avg_load) * local->group_power
5832 ) / SCHED_POWER_SCALE;
5835 * if *imbalance is less than the average load per runnable task
5836 * there is no guarantee that any tasks will be moved so we'll have
5837 * a think about bumping its value to force at least one task to be
5840 if (env->imbalance < busiest->load_per_task)
5841 return fix_small_imbalance(env, sds);
5844 /******* find_busiest_group() helpers end here *********************/
5847 * find_busiest_group - Returns the busiest group within the sched_domain
5848 * if there is an imbalance. If there isn't an imbalance, and
5849 * the user has opted for power-savings, it returns a group whose
5850 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5851 * such a group exists.
5853 * Also calculates the amount of weighted load which should be moved
5854 * to restore balance.
5856 * @env: The load balancing environment.
5858 * Return: - The busiest group if imbalance exists.
5859 * - If no imbalance and user has opted for power-savings balance,
5860 * return the least loaded group whose CPUs can be
5861 * put to idle by rebalancing its tasks onto our group.
5863 static struct sched_group *find_busiest_group(struct lb_env *env)
5865 struct sg_lb_stats *local, *busiest;
5866 struct sd_lb_stats sds;
5868 init_sd_lb_stats(&sds);
5871 * Compute the various statistics relavent for load balancing at
5874 update_sd_lb_stats(env, &sds);
5875 local = &sds.local_stat;
5876 busiest = &sds.busiest_stat;
5878 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5879 check_asym_packing(env, &sds))
5882 /* There is no busy sibling group to pull tasks from */
5883 if (!sds.busiest || busiest->sum_nr_running == 0)
5886 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5889 * If the busiest group is imbalanced the below checks don't
5890 * work because they assume all things are equal, which typically
5891 * isn't true due to cpus_allowed constraints and the like.
5893 if (busiest->group_imb)
5896 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5897 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5898 !busiest->group_has_capacity)
5902 * If the local group is more busy than the selected busiest group
5903 * don't try and pull any tasks.
5905 if (local->avg_load >= busiest->avg_load)
5909 * Don't pull any tasks if this group is already above the domain
5912 if (local->avg_load >= sds.avg_load)
5915 if (env->idle == CPU_IDLE) {
5917 * This cpu is idle. If the busiest group load doesn't
5918 * have more tasks than the number of available cpu's and
5919 * there is no imbalance between this and busiest group
5920 * wrt to idle cpu's, it is balanced.
5922 if ((local->idle_cpus < busiest->idle_cpus) &&
5923 busiest->sum_nr_running <= busiest->group_weight)
5927 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5928 * imbalance_pct to be conservative.
5930 if (100 * busiest->avg_load <=
5931 env->sd->imbalance_pct * local->avg_load)
5936 /* Looks like there is an imbalance. Compute it */
5937 calculate_imbalance(env, &sds);
5946 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5948 static struct rq *find_busiest_queue(struct lb_env *env,
5949 struct sched_group *group)
5951 struct rq *busiest = NULL, *rq;
5952 unsigned long busiest_load = 0, busiest_power = 1;
5955 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5956 unsigned long power, capacity, wl;
5960 rt = fbq_classify_rq(rq);
5963 * We classify groups/runqueues into three groups:
5964 * - regular: there are !numa tasks
5965 * - remote: there are numa tasks that run on the 'wrong' node
5966 * - all: there is no distinction
5968 * In order to avoid migrating ideally placed numa tasks,
5969 * ignore those when there's better options.
5971 * If we ignore the actual busiest queue to migrate another
5972 * task, the next balance pass can still reduce the busiest
5973 * queue by moving tasks around inside the node.
5975 * If we cannot move enough load due to this classification
5976 * the next pass will adjust the group classification and
5977 * allow migration of more tasks.
5979 * Both cases only affect the total convergence complexity.
5981 if (rt > env->fbq_type)
5984 power = power_of(i);
5985 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5987 capacity = fix_small_capacity(env->sd, group);
5989 wl = weighted_cpuload(i);
5992 * When comparing with imbalance, use weighted_cpuload()
5993 * which is not scaled with the cpu power.
5995 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5999 * For the load comparisons with the other cpu's, consider
6000 * the weighted_cpuload() scaled with the cpu power, so that
6001 * the load can be moved away from the cpu that is potentially
6002 * running at a lower capacity.
6004 * Thus we're looking for max(wl_i / power_i), crosswise
6005 * multiplication to rid ourselves of the division works out
6006 * to: wl_i * power_j > wl_j * power_i; where j is our
6009 if (wl * busiest_power > busiest_load * power) {
6011 busiest_power = power;
6020 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6021 * so long as it is large enough.
6023 #define MAX_PINNED_INTERVAL 512
6025 /* Working cpumask for load_balance and load_balance_newidle. */
6026 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6028 static int need_active_balance(struct lb_env *env)
6030 struct sched_domain *sd = env->sd;
6032 if (env->idle == CPU_NEWLY_IDLE) {
6035 * ASYM_PACKING needs to force migrate tasks from busy but
6036 * higher numbered CPUs in order to pack all tasks in the
6037 * lowest numbered CPUs.
6039 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6043 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6046 static int active_load_balance_cpu_stop(void *data);
6048 static int should_we_balance(struct lb_env *env)
6050 struct sched_group *sg = env->sd->groups;
6051 struct cpumask *sg_cpus, *sg_mask;
6052 int cpu, balance_cpu = -1;
6055 * In the newly idle case, we will allow all the cpu's
6056 * to do the newly idle load balance.
6058 if (env->idle == CPU_NEWLY_IDLE)
6061 sg_cpus = sched_group_cpus(sg);
6062 sg_mask = sched_group_mask(sg);
6063 /* Try to find first idle cpu */
6064 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6065 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6072 if (balance_cpu == -1)
6073 balance_cpu = group_balance_cpu(sg);
6076 * First idle cpu or the first cpu(busiest) in this sched group
6077 * is eligible for doing load balancing at this and above domains.
6079 return balance_cpu == env->dst_cpu;
6083 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6084 * tasks if there is an imbalance.
6086 static int load_balance(int this_cpu, struct rq *this_rq,
6087 struct sched_domain *sd, enum cpu_idle_type idle,
6088 int *continue_balancing)
6090 int ld_moved, cur_ld_moved, active_balance = 0;
6091 struct sched_domain *sd_parent = sd->parent;
6092 struct sched_group *group;
6094 unsigned long flags;
6095 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6097 struct lb_env env = {
6099 .dst_cpu = this_cpu,
6101 .dst_grpmask = sched_group_cpus(sd->groups),
6103 .loop_break = sched_nr_migrate_break,
6109 * For NEWLY_IDLE load_balancing, we don't need to consider
6110 * other cpus in our group
6112 if (idle == CPU_NEWLY_IDLE)
6113 env.dst_grpmask = NULL;
6115 cpumask_copy(cpus, cpu_active_mask);
6117 schedstat_inc(sd, lb_count[idle]);
6120 if (!should_we_balance(&env)) {
6121 *continue_balancing = 0;
6125 group = find_busiest_group(&env);
6127 schedstat_inc(sd, lb_nobusyg[idle]);
6131 busiest = find_busiest_queue(&env, group);
6133 schedstat_inc(sd, lb_nobusyq[idle]);
6137 BUG_ON(busiest == env.dst_rq);
6139 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6142 if (busiest->nr_running > 1) {
6144 * Attempt to move tasks. If find_busiest_group has found
6145 * an imbalance but busiest->nr_running <= 1, the group is
6146 * still unbalanced. ld_moved simply stays zero, so it is
6147 * correctly treated as an imbalance.
6149 env.flags |= LBF_ALL_PINNED;
6150 env.src_cpu = busiest->cpu;
6151 env.src_rq = busiest;
6152 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6155 local_irq_save(flags);
6156 double_rq_lock(env.dst_rq, busiest);
6159 * cur_ld_moved - load moved in current iteration
6160 * ld_moved - cumulative load moved across iterations
6162 cur_ld_moved = move_tasks(&env);
6163 ld_moved += cur_ld_moved;
6164 double_rq_unlock(env.dst_rq, busiest);
6165 local_irq_restore(flags);
6168 * some other cpu did the load balance for us.
6170 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6171 resched_cpu(env.dst_cpu);
6173 if (env.flags & LBF_NEED_BREAK) {
6174 env.flags &= ~LBF_NEED_BREAK;
6179 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6180 * us and move them to an alternate dst_cpu in our sched_group
6181 * where they can run. The upper limit on how many times we
6182 * iterate on same src_cpu is dependent on number of cpus in our
6185 * This changes load balance semantics a bit on who can move
6186 * load to a given_cpu. In addition to the given_cpu itself
6187 * (or a ilb_cpu acting on its behalf where given_cpu is
6188 * nohz-idle), we now have balance_cpu in a position to move
6189 * load to given_cpu. In rare situations, this may cause
6190 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6191 * _independently_ and at _same_ time to move some load to
6192 * given_cpu) causing exceess load to be moved to given_cpu.
6193 * This however should not happen so much in practice and
6194 * moreover subsequent load balance cycles should correct the
6195 * excess load moved.
6197 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6199 /* Prevent to re-select dst_cpu via env's cpus */
6200 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6202 env.dst_rq = cpu_rq(env.new_dst_cpu);
6203 env.dst_cpu = env.new_dst_cpu;
6204 env.flags &= ~LBF_DST_PINNED;
6206 env.loop_break = sched_nr_migrate_break;
6209 * Go back to "more_balance" rather than "redo" since we
6210 * need to continue with same src_cpu.
6216 * We failed to reach balance because of affinity.
6219 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6221 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6222 *group_imbalance = 1;
6223 } else if (*group_imbalance)
6224 *group_imbalance = 0;
6227 /* All tasks on this runqueue were pinned by CPU affinity */
6228 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6229 cpumask_clear_cpu(cpu_of(busiest), cpus);
6230 if (!cpumask_empty(cpus)) {
6232 env.loop_break = sched_nr_migrate_break;
6240 schedstat_inc(sd, lb_failed[idle]);
6242 * Increment the failure counter only on periodic balance.
6243 * We do not want newidle balance, which can be very
6244 * frequent, pollute the failure counter causing
6245 * excessive cache_hot migrations and active balances.
6247 if (idle != CPU_NEWLY_IDLE)
6248 sd->nr_balance_failed++;
6250 if (need_active_balance(&env)) {
6251 raw_spin_lock_irqsave(&busiest->lock, flags);
6253 /* don't kick the active_load_balance_cpu_stop,
6254 * if the curr task on busiest cpu can't be
6257 if (!cpumask_test_cpu(this_cpu,
6258 tsk_cpus_allowed(busiest->curr))) {
6259 raw_spin_unlock_irqrestore(&busiest->lock,
6261 env.flags |= LBF_ALL_PINNED;
6262 goto out_one_pinned;
6266 * ->active_balance synchronizes accesses to
6267 * ->active_balance_work. Once set, it's cleared
6268 * only after active load balance is finished.
6270 if (!busiest->active_balance) {
6271 busiest->active_balance = 1;
6272 busiest->push_cpu = this_cpu;
6275 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6277 if (active_balance) {
6278 stop_one_cpu_nowait(cpu_of(busiest),
6279 active_load_balance_cpu_stop, busiest,
6280 &busiest->active_balance_work);
6284 * We've kicked active balancing, reset the failure
6287 sd->nr_balance_failed = sd->cache_nice_tries+1;
6290 sd->nr_balance_failed = 0;
6292 if (likely(!active_balance)) {
6293 /* We were unbalanced, so reset the balancing interval */
6294 sd->balance_interval = sd->min_interval;
6297 * If we've begun active balancing, start to back off. This
6298 * case may not be covered by the all_pinned logic if there
6299 * is only 1 task on the busy runqueue (because we don't call
6302 if (sd->balance_interval < sd->max_interval)
6303 sd->balance_interval *= 2;
6309 schedstat_inc(sd, lb_balanced[idle]);
6311 sd->nr_balance_failed = 0;
6314 /* tune up the balancing interval */
6315 if (((env.flags & LBF_ALL_PINNED) &&
6316 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6317 (sd->balance_interval < sd->max_interval))
6318 sd->balance_interval *= 2;
6326 * idle_balance is called by schedule() if this_cpu is about to become
6327 * idle. Attempts to pull tasks from other CPUs.
6329 void idle_balance(int this_cpu, struct rq *this_rq)
6331 struct sched_domain *sd;
6332 int pulled_task = 0;
6333 unsigned long next_balance = jiffies + HZ;
6336 this_rq->idle_stamp = rq_clock(this_rq);
6338 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6342 * Drop the rq->lock, but keep IRQ/preempt disabled.
6344 raw_spin_unlock(&this_rq->lock);
6346 update_blocked_averages(this_cpu);
6348 for_each_domain(this_cpu, sd) {
6349 unsigned long interval;
6350 int continue_balancing = 1;
6351 u64 t0, domain_cost;
6353 if (!(sd->flags & SD_LOAD_BALANCE))
6356 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6359 if (sd->flags & SD_BALANCE_NEWIDLE) {
6360 t0 = sched_clock_cpu(this_cpu);
6362 /* If we've pulled tasks over stop searching: */
6363 pulled_task = load_balance(this_cpu, this_rq,
6365 &continue_balancing);
6367 domain_cost = sched_clock_cpu(this_cpu) - t0;
6368 if (domain_cost > sd->max_newidle_lb_cost)
6369 sd->max_newidle_lb_cost = domain_cost;
6371 curr_cost += domain_cost;
6374 interval = msecs_to_jiffies(sd->balance_interval);
6375 if (time_after(next_balance, sd->last_balance + interval))
6376 next_balance = sd->last_balance + interval;
6378 this_rq->idle_stamp = 0;
6384 raw_spin_lock(&this_rq->lock);
6386 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6388 * We are going idle. next_balance may be set based on
6389 * a busy processor. So reset next_balance.
6391 this_rq->next_balance = next_balance;
6394 if (curr_cost > this_rq->max_idle_balance_cost)
6395 this_rq->max_idle_balance_cost = curr_cost;
6399 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6400 * running tasks off the busiest CPU onto idle CPUs. It requires at
6401 * least 1 task to be running on each physical CPU where possible, and
6402 * avoids physical / logical imbalances.
6404 static int active_load_balance_cpu_stop(void *data)
6406 struct rq *busiest_rq = data;
6407 int busiest_cpu = cpu_of(busiest_rq);
6408 int target_cpu = busiest_rq->push_cpu;
6409 struct rq *target_rq = cpu_rq(target_cpu);
6410 struct sched_domain *sd;
6412 raw_spin_lock_irq(&busiest_rq->lock);
6414 /* make sure the requested cpu hasn't gone down in the meantime */
6415 if (unlikely(busiest_cpu != smp_processor_id() ||
6416 !busiest_rq->active_balance))
6419 /* Is there any task to move? */
6420 if (busiest_rq->nr_running <= 1)
6424 * This condition is "impossible", if it occurs
6425 * we need to fix it. Originally reported by
6426 * Bjorn Helgaas on a 128-cpu setup.
6428 BUG_ON(busiest_rq == target_rq);
6430 /* move a task from busiest_rq to target_rq */
6431 double_lock_balance(busiest_rq, target_rq);
6433 /* Search for an sd spanning us and the target CPU. */
6435 for_each_domain(target_cpu, sd) {
6436 if ((sd->flags & SD_LOAD_BALANCE) &&
6437 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6442 struct lb_env env = {
6444 .dst_cpu = target_cpu,
6445 .dst_rq = target_rq,
6446 .src_cpu = busiest_rq->cpu,
6447 .src_rq = busiest_rq,
6451 schedstat_inc(sd, alb_count);
6453 if (move_one_task(&env))
6454 schedstat_inc(sd, alb_pushed);
6456 schedstat_inc(sd, alb_failed);
6459 double_unlock_balance(busiest_rq, target_rq);
6461 busiest_rq->active_balance = 0;
6462 raw_spin_unlock_irq(&busiest_rq->lock);
6466 #ifdef CONFIG_NO_HZ_COMMON
6468 * idle load balancing details
6469 * - When one of the busy CPUs notice that there may be an idle rebalancing
6470 * needed, they will kick the idle load balancer, which then does idle
6471 * load balancing for all the idle CPUs.
6474 cpumask_var_t idle_cpus_mask;
6476 unsigned long next_balance; /* in jiffy units */
6477 } nohz ____cacheline_aligned;
6479 static inline int find_new_ilb(int call_cpu)
6481 int ilb = cpumask_first(nohz.idle_cpus_mask);
6483 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6490 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6491 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6492 * CPU (if there is one).
6494 static void nohz_balancer_kick(int cpu)
6498 nohz.next_balance++;
6500 ilb_cpu = find_new_ilb(cpu);
6502 if (ilb_cpu >= nr_cpu_ids)
6505 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6508 * Use smp_send_reschedule() instead of resched_cpu().
6509 * This way we generate a sched IPI on the target cpu which
6510 * is idle. And the softirq performing nohz idle load balance
6511 * will be run before returning from the IPI.
6513 smp_send_reschedule(ilb_cpu);
6517 static inline void nohz_balance_exit_idle(int cpu)
6519 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6520 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6521 atomic_dec(&nohz.nr_cpus);
6522 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6526 static inline void set_cpu_sd_state_busy(void)
6528 struct sched_domain *sd;
6531 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6533 if (!sd || !sd->nohz_idle)
6537 for (; sd; sd = sd->parent)
6538 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6543 void set_cpu_sd_state_idle(void)
6545 struct sched_domain *sd;
6548 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6550 if (!sd || sd->nohz_idle)
6554 for (; sd; sd = sd->parent)
6555 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6561 * This routine will record that the cpu is going idle with tick stopped.
6562 * This info will be used in performing idle load balancing in the future.
6564 void nohz_balance_enter_idle(int cpu)
6567 * If this cpu is going down, then nothing needs to be done.
6569 if (!cpu_active(cpu))
6572 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6575 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6576 atomic_inc(&nohz.nr_cpus);
6577 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6580 static int sched_ilb_notifier(struct notifier_block *nfb,
6581 unsigned long action, void *hcpu)
6583 switch (action & ~CPU_TASKS_FROZEN) {
6585 nohz_balance_exit_idle(smp_processor_id());
6593 static DEFINE_SPINLOCK(balancing);
6596 * Scale the max load_balance interval with the number of CPUs in the system.
6597 * This trades load-balance latency on larger machines for less cross talk.
6599 void update_max_interval(void)
6601 max_load_balance_interval = HZ*num_online_cpus()/10;
6605 * It checks each scheduling domain to see if it is due to be balanced,
6606 * and initiates a balancing operation if so.
6608 * Balancing parameters are set up in init_sched_domains.
6610 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6612 int continue_balancing = 1;
6613 struct rq *rq = cpu_rq(cpu);
6614 unsigned long interval;
6615 struct sched_domain *sd;
6616 /* Earliest time when we have to do rebalance again */
6617 unsigned long next_balance = jiffies + 60*HZ;
6618 int update_next_balance = 0;
6619 int need_serialize, need_decay = 0;
6622 update_blocked_averages(cpu);
6625 for_each_domain(cpu, sd) {
6627 * Decay the newidle max times here because this is a regular
6628 * visit to all the domains. Decay ~1% per second.
6630 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6631 sd->max_newidle_lb_cost =
6632 (sd->max_newidle_lb_cost * 253) / 256;
6633 sd->next_decay_max_lb_cost = jiffies + HZ;
6636 max_cost += sd->max_newidle_lb_cost;
6638 if (!(sd->flags & SD_LOAD_BALANCE))
6642 * Stop the load balance at this level. There is another
6643 * CPU in our sched group which is doing load balancing more
6646 if (!continue_balancing) {
6652 interval = sd->balance_interval;
6653 if (idle != CPU_IDLE)
6654 interval *= sd->busy_factor;
6656 /* scale ms to jiffies */
6657 interval = msecs_to_jiffies(interval);
6658 interval = clamp(interval, 1UL, max_load_balance_interval);
6660 need_serialize = sd->flags & SD_SERIALIZE;
6662 if (need_serialize) {
6663 if (!spin_trylock(&balancing))
6667 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6668 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6670 * The LBF_DST_PINNED logic could have changed
6671 * env->dst_cpu, so we can't know our idle
6672 * state even if we migrated tasks. Update it.
6674 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6676 sd->last_balance = jiffies;
6679 spin_unlock(&balancing);
6681 if (time_after(next_balance, sd->last_balance + interval)) {
6682 next_balance = sd->last_balance + interval;
6683 update_next_balance = 1;
6688 * Ensure the rq-wide value also decays but keep it at a
6689 * reasonable floor to avoid funnies with rq->avg_idle.
6691 rq->max_idle_balance_cost =
6692 max((u64)sysctl_sched_migration_cost, max_cost);
6697 * next_balance will be updated only when there is a need.
6698 * When the cpu is attached to null domain for ex, it will not be
6701 if (likely(update_next_balance))
6702 rq->next_balance = next_balance;
6705 #ifdef CONFIG_NO_HZ_COMMON
6707 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6708 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6710 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6712 struct rq *this_rq = cpu_rq(this_cpu);
6716 if (idle != CPU_IDLE ||
6717 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6720 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6721 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6725 * If this cpu gets work to do, stop the load balancing
6726 * work being done for other cpus. Next load
6727 * balancing owner will pick it up.
6732 rq = cpu_rq(balance_cpu);
6734 raw_spin_lock_irq(&rq->lock);
6735 update_rq_clock(rq);
6736 update_idle_cpu_load(rq);
6737 raw_spin_unlock_irq(&rq->lock);
6739 rebalance_domains(balance_cpu, CPU_IDLE);
6741 if (time_after(this_rq->next_balance, rq->next_balance))
6742 this_rq->next_balance = rq->next_balance;
6744 nohz.next_balance = this_rq->next_balance;
6746 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6750 * Current heuristic for kicking the idle load balancer in the presence
6751 * of an idle cpu is the system.
6752 * - This rq has more than one task.
6753 * - At any scheduler domain level, this cpu's scheduler group has multiple
6754 * busy cpu's exceeding the group's power.
6755 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6756 * domain span are idle.
6758 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6760 unsigned long now = jiffies;
6761 struct sched_domain *sd;
6763 if (unlikely(idle_cpu(cpu)))
6767 * We may be recently in ticked or tickless idle mode. At the first
6768 * busy tick after returning from idle, we will update the busy stats.
6770 set_cpu_sd_state_busy();
6771 nohz_balance_exit_idle(cpu);
6774 * None are in tickless mode and hence no need for NOHZ idle load
6777 if (likely(!atomic_read(&nohz.nr_cpus)))
6780 if (time_before(now, nohz.next_balance))
6783 if (rq->nr_running >= 2)
6787 for_each_domain(cpu, sd) {
6788 struct sched_group *sg = sd->groups;
6789 struct sched_group_power *sgp = sg->sgp;
6790 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6792 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6793 goto need_kick_unlock;
6795 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6796 && (cpumask_first_and(nohz.idle_cpus_mask,
6797 sched_domain_span(sd)) < cpu))
6798 goto need_kick_unlock;
6800 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6812 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6816 * run_rebalance_domains is triggered when needed from the scheduler tick.
6817 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6819 static void run_rebalance_domains(struct softirq_action *h)
6821 int this_cpu = smp_processor_id();
6822 struct rq *this_rq = cpu_rq(this_cpu);
6823 enum cpu_idle_type idle = this_rq->idle_balance ?
6824 CPU_IDLE : CPU_NOT_IDLE;
6826 rebalance_domains(this_cpu, idle);
6829 * If this cpu has a pending nohz_balance_kick, then do the
6830 * balancing on behalf of the other idle cpus whose ticks are
6833 nohz_idle_balance(this_cpu, idle);
6836 static inline int on_null_domain(int cpu)
6838 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6842 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6844 void trigger_load_balance(struct rq *rq, int cpu)
6846 /* Don't need to rebalance while attached to NULL domain */
6847 if (time_after_eq(jiffies, rq->next_balance) &&
6848 likely(!on_null_domain(cpu)))
6849 raise_softirq(SCHED_SOFTIRQ);
6850 #ifdef CONFIG_NO_HZ_COMMON
6851 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6852 nohz_balancer_kick(cpu);
6856 static void rq_online_fair(struct rq *rq)
6861 static void rq_offline_fair(struct rq *rq)
6865 /* Ensure any throttled groups are reachable by pick_next_task */
6866 unthrottle_offline_cfs_rqs(rq);
6869 #endif /* CONFIG_SMP */
6872 * scheduler tick hitting a task of our scheduling class:
6874 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6876 struct cfs_rq *cfs_rq;
6877 struct sched_entity *se = &curr->se;
6879 for_each_sched_entity(se) {
6880 cfs_rq = cfs_rq_of(se);
6881 entity_tick(cfs_rq, se, queued);
6884 if (numabalancing_enabled)
6885 task_tick_numa(rq, curr);
6887 update_rq_runnable_avg(rq, 1);
6891 * called on fork with the child task as argument from the parent's context
6892 * - child not yet on the tasklist
6893 * - preemption disabled
6895 static void task_fork_fair(struct task_struct *p)
6897 struct cfs_rq *cfs_rq;
6898 struct sched_entity *se = &p->se, *curr;
6899 int this_cpu = smp_processor_id();
6900 struct rq *rq = this_rq();
6901 unsigned long flags;
6903 raw_spin_lock_irqsave(&rq->lock, flags);
6905 update_rq_clock(rq);
6907 cfs_rq = task_cfs_rq(current);
6908 curr = cfs_rq->curr;
6911 * Not only the cpu but also the task_group of the parent might have
6912 * been changed after parent->se.parent,cfs_rq were copied to
6913 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6914 * of child point to valid ones.
6917 __set_task_cpu(p, this_cpu);
6920 update_curr(cfs_rq);
6923 se->vruntime = curr->vruntime;
6924 place_entity(cfs_rq, se, 1);
6926 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6928 * Upon rescheduling, sched_class::put_prev_task() will place
6929 * 'current' within the tree based on its new key value.
6931 swap(curr->vruntime, se->vruntime);
6932 resched_task(rq->curr);
6935 se->vruntime -= cfs_rq->min_vruntime;
6937 raw_spin_unlock_irqrestore(&rq->lock, flags);
6941 * Priority of the task has changed. Check to see if we preempt
6945 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6951 * Reschedule if we are currently running on this runqueue and
6952 * our priority decreased, or if we are not currently running on
6953 * this runqueue and our priority is higher than the current's
6955 if (rq->curr == p) {
6956 if (p->prio > oldprio)
6957 resched_task(rq->curr);
6959 check_preempt_curr(rq, p, 0);
6962 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6964 struct sched_entity *se = &p->se;
6965 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6968 * Ensure the task's vruntime is normalized, so that when its
6969 * switched back to the fair class the enqueue_entity(.flags=0) will
6970 * do the right thing.
6972 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6973 * have normalized the vruntime, if it was !on_rq, then only when
6974 * the task is sleeping will it still have non-normalized vruntime.
6976 if (!se->on_rq && p->state != TASK_RUNNING) {
6978 * Fix up our vruntime so that the current sleep doesn't
6979 * cause 'unlimited' sleep bonus.
6981 place_entity(cfs_rq, se, 0);
6982 se->vruntime -= cfs_rq->min_vruntime;
6987 * Remove our load from contribution when we leave sched_fair
6988 * and ensure we don't carry in an old decay_count if we
6991 if (se->avg.decay_count) {
6992 __synchronize_entity_decay(se);
6993 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6999 * We switched to the sched_fair class.
7001 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7007 * We were most likely switched from sched_rt, so
7008 * kick off the schedule if running, otherwise just see
7009 * if we can still preempt the current task.
7012 resched_task(rq->curr);
7014 check_preempt_curr(rq, p, 0);
7017 /* Account for a task changing its policy or group.
7019 * This routine is mostly called to set cfs_rq->curr field when a task
7020 * migrates between groups/classes.
7022 static void set_curr_task_fair(struct rq *rq)
7024 struct sched_entity *se = &rq->curr->se;
7026 for_each_sched_entity(se) {
7027 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7029 set_next_entity(cfs_rq, se);
7030 /* ensure bandwidth has been allocated on our new cfs_rq */
7031 account_cfs_rq_runtime(cfs_rq, 0);
7035 void init_cfs_rq(struct cfs_rq *cfs_rq)
7037 cfs_rq->tasks_timeline = RB_ROOT;
7038 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7039 #ifndef CONFIG_64BIT
7040 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7043 atomic64_set(&cfs_rq->decay_counter, 1);
7044 atomic_long_set(&cfs_rq->removed_load, 0);
7048 #ifdef CONFIG_FAIR_GROUP_SCHED
7049 static void task_move_group_fair(struct task_struct *p, int on_rq)
7051 struct cfs_rq *cfs_rq;
7053 * If the task was not on the rq at the time of this cgroup movement
7054 * it must have been asleep, sleeping tasks keep their ->vruntime
7055 * absolute on their old rq until wakeup (needed for the fair sleeper
7056 * bonus in place_entity()).
7058 * If it was on the rq, we've just 'preempted' it, which does convert
7059 * ->vruntime to a relative base.
7061 * Make sure both cases convert their relative position when migrating
7062 * to another cgroup's rq. This does somewhat interfere with the
7063 * fair sleeper stuff for the first placement, but who cares.
7066 * When !on_rq, vruntime of the task has usually NOT been normalized.
7067 * But there are some cases where it has already been normalized:
7069 * - Moving a forked child which is waiting for being woken up by
7070 * wake_up_new_task().
7071 * - Moving a task which has been woken up by try_to_wake_up() and
7072 * waiting for actually being woken up by sched_ttwu_pending().
7074 * To prevent boost or penalty in the new cfs_rq caused by delta
7075 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7077 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7081 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7082 set_task_rq(p, task_cpu(p));
7084 cfs_rq = cfs_rq_of(&p->se);
7085 p->se.vruntime += cfs_rq->min_vruntime;
7088 * migrate_task_rq_fair() will have removed our previous
7089 * contribution, but we must synchronize for ongoing future
7092 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7093 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7098 void free_fair_sched_group(struct task_group *tg)
7102 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7104 for_each_possible_cpu(i) {
7106 kfree(tg->cfs_rq[i]);
7115 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7117 struct cfs_rq *cfs_rq;
7118 struct sched_entity *se;
7121 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7124 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7128 tg->shares = NICE_0_LOAD;
7130 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7132 for_each_possible_cpu(i) {
7133 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7134 GFP_KERNEL, cpu_to_node(i));
7138 se = kzalloc_node(sizeof(struct sched_entity),
7139 GFP_KERNEL, cpu_to_node(i));
7143 init_cfs_rq(cfs_rq);
7144 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7155 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7157 struct rq *rq = cpu_rq(cpu);
7158 unsigned long flags;
7161 * Only empty task groups can be destroyed; so we can speculatively
7162 * check on_list without danger of it being re-added.
7164 if (!tg->cfs_rq[cpu]->on_list)
7167 raw_spin_lock_irqsave(&rq->lock, flags);
7168 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7169 raw_spin_unlock_irqrestore(&rq->lock, flags);
7172 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7173 struct sched_entity *se, int cpu,
7174 struct sched_entity *parent)
7176 struct rq *rq = cpu_rq(cpu);
7180 init_cfs_rq_runtime(cfs_rq);
7182 tg->cfs_rq[cpu] = cfs_rq;
7185 /* se could be NULL for root_task_group */
7190 se->cfs_rq = &rq->cfs;
7192 se->cfs_rq = parent->my_q;
7195 update_load_set(&se->load, 0);
7196 se->parent = parent;
7199 static DEFINE_MUTEX(shares_mutex);
7201 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7204 unsigned long flags;
7207 * We can't change the weight of the root cgroup.
7212 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7214 mutex_lock(&shares_mutex);
7215 if (tg->shares == shares)
7218 tg->shares = shares;
7219 for_each_possible_cpu(i) {
7220 struct rq *rq = cpu_rq(i);
7221 struct sched_entity *se;
7224 /* Propagate contribution to hierarchy */
7225 raw_spin_lock_irqsave(&rq->lock, flags);
7227 /* Possible calls to update_curr() need rq clock */
7228 update_rq_clock(rq);
7229 for_each_sched_entity(se)
7230 update_cfs_shares(group_cfs_rq(se));
7231 raw_spin_unlock_irqrestore(&rq->lock, flags);
7235 mutex_unlock(&shares_mutex);
7238 #else /* CONFIG_FAIR_GROUP_SCHED */
7240 void free_fair_sched_group(struct task_group *tg) { }
7242 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7247 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7249 #endif /* CONFIG_FAIR_GROUP_SCHED */
7252 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7254 struct sched_entity *se = &task->se;
7255 unsigned int rr_interval = 0;
7258 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7261 if (rq->cfs.load.weight)
7262 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7268 * All the scheduling class methods:
7270 const struct sched_class fair_sched_class = {
7271 .next = &idle_sched_class,
7272 .enqueue_task = enqueue_task_fair,
7273 .dequeue_task = dequeue_task_fair,
7274 .yield_task = yield_task_fair,
7275 .yield_to_task = yield_to_task_fair,
7277 .check_preempt_curr = check_preempt_wakeup,
7279 .pick_next_task = pick_next_task_fair,
7280 .put_prev_task = put_prev_task_fair,
7283 .select_task_rq = select_task_rq_fair,
7284 .migrate_task_rq = migrate_task_rq_fair,
7286 .rq_online = rq_online_fair,
7287 .rq_offline = rq_offline_fair,
7289 .task_waking = task_waking_fair,
7292 .set_curr_task = set_curr_task_fair,
7293 .task_tick = task_tick_fair,
7294 .task_fork = task_fork_fair,
7296 .prio_changed = prio_changed_fair,
7297 .switched_from = switched_from_fair,
7298 .switched_to = switched_to_fair,
7300 .get_rr_interval = get_rr_interval_fair,
7302 #ifdef CONFIG_FAIR_GROUP_SCHED
7303 .task_move_group = task_move_group_fair,
7307 #ifdef CONFIG_SCHED_DEBUG
7308 void print_cfs_stats(struct seq_file *m, int cpu)
7310 struct cfs_rq *cfs_rq;
7313 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7314 print_cfs_rq(m, cpu, cfs_rq);
7319 __init void init_sched_fair_class(void)
7322 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7324 #ifdef CONFIG_NO_HZ_COMMON
7325 nohz.next_balance = jiffies;
7326 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7327 cpu_notifier(sched_ilb_notifier, 0);