2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
830 /* Portion of address space to scan in MB */
831 unsigned int sysctl_numa_balancing_scan_size = 256;
833 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
834 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 * After skipping a page migration on a shared page, skip N more numa page
838 * migrations unconditionally. This reduces the number of NUMA migrations
839 * in shared memory workloads, and has the effect of pulling tasks towards
840 * where their memory lives, over pulling the memory towards the task.
842 unsigned int sysctl_numa_balancing_migrate_deferred = 16;
844 static unsigned int task_nr_scan_windows(struct task_struct *p)
846 unsigned long rss = 0;
847 unsigned long nr_scan_pages;
850 * Calculations based on RSS as non-present and empty pages are skipped
851 * by the PTE scanner and NUMA hinting faults should be trapped based
854 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
855 rss = get_mm_rss(p->mm);
859 rss = round_up(rss, nr_scan_pages);
860 return rss / nr_scan_pages;
863 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
864 #define MAX_SCAN_WINDOW 2560
866 static unsigned int task_scan_min(struct task_struct *p)
868 unsigned int scan, floor;
869 unsigned int windows = 1;
871 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
872 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
873 floor = 1000 / windows;
875 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
876 return max_t(unsigned int, floor, scan);
879 static unsigned int task_scan_max(struct task_struct *p)
881 unsigned int smin = task_scan_min(p);
884 /* Watch for min being lower than max due to floor calculations */
885 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
886 return max(smin, smax);
890 * Once a preferred node is selected the scheduler balancer will prefer moving
891 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
892 * scans. This will give the process the chance to accumulate more faults on
893 * the preferred node but still allow the scheduler to move the task again if
894 * the nodes CPUs are overloaded.
896 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
898 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
900 rq->nr_numa_running += (p->numa_preferred_nid != -1);
901 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
904 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
906 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
907 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
913 spinlock_t lock; /* nr_tasks, tasks */
916 struct list_head task_list;
919 unsigned long total_faults;
920 unsigned long faults[0];
923 pid_t task_numa_group_id(struct task_struct *p)
925 return p->numa_group ? p->numa_group->gid : 0;
928 static inline int task_faults_idx(int nid, int priv)
930 return 2 * nid + priv;
933 static inline unsigned long task_faults(struct task_struct *p, int nid)
938 return p->numa_faults[task_faults_idx(nid, 0)] +
939 p->numa_faults[task_faults_idx(nid, 1)];
942 static inline unsigned long group_faults(struct task_struct *p, int nid)
947 return p->numa_group->faults[2*nid] + p->numa_group->faults[2*nid+1];
951 * These return the fraction of accesses done by a particular task, or
952 * task group, on a particular numa node. The group weight is given a
953 * larger multiplier, in order to group tasks together that are almost
954 * evenly spread out between numa nodes.
956 static inline unsigned long task_weight(struct task_struct *p, int nid)
958 unsigned long total_faults;
963 total_faults = p->total_numa_faults;
968 return 1000 * task_faults(p, nid) / total_faults;
971 static inline unsigned long group_weight(struct task_struct *p, int nid)
973 if (!p->numa_group || !p->numa_group->total_faults)
976 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
979 static unsigned long weighted_cpuload(const int cpu);
980 static unsigned long source_load(int cpu, int type);
981 static unsigned long target_load(int cpu, int type);
982 static unsigned long power_of(int cpu);
983 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
985 /* Cached statistics for all CPUs within a node */
987 unsigned long nr_running;
990 /* Total compute capacity of CPUs on a node */
993 /* Approximate capacity in terms of runnable tasks on a node */
994 unsigned long capacity;
999 * XXX borrowed from update_sg_lb_stats
1001 static void update_numa_stats(struct numa_stats *ns, int nid)
1005 memset(ns, 0, sizeof(*ns));
1006 for_each_cpu(cpu, cpumask_of_node(nid)) {
1007 struct rq *rq = cpu_rq(cpu);
1009 ns->nr_running += rq->nr_running;
1010 ns->load += weighted_cpuload(cpu);
1011 ns->power += power_of(cpu);
1014 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1015 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1016 ns->has_capacity = (ns->nr_running < ns->capacity);
1019 struct task_numa_env {
1020 struct task_struct *p;
1022 int src_cpu, src_nid;
1023 int dst_cpu, dst_nid;
1025 struct numa_stats src_stats, dst_stats;
1027 int imbalance_pct, idx;
1029 struct task_struct *best_task;
1034 static void task_numa_assign(struct task_numa_env *env,
1035 struct task_struct *p, long imp)
1038 put_task_struct(env->best_task);
1043 env->best_imp = imp;
1044 env->best_cpu = env->dst_cpu;
1048 * This checks if the overall compute and NUMA accesses of the system would
1049 * be improved if the source tasks was migrated to the target dst_cpu taking
1050 * into account that it might be best if task running on the dst_cpu should
1051 * be exchanged with the source task
1053 static void task_numa_compare(struct task_numa_env *env,
1054 long taskimp, long groupimp)
1056 struct rq *src_rq = cpu_rq(env->src_cpu);
1057 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1058 struct task_struct *cur;
1059 long dst_load, src_load;
1061 long imp = (groupimp > 0) ? groupimp : taskimp;
1064 cur = ACCESS_ONCE(dst_rq->curr);
1065 if (cur->pid == 0) /* idle */
1069 * "imp" is the fault differential for the source task between the
1070 * source and destination node. Calculate the total differential for
1071 * the source task and potential destination task. The more negative
1072 * the value is, the more rmeote accesses that would be expected to
1073 * be incurred if the tasks were swapped.
1076 /* Skip this swap candidate if cannot move to the source cpu */
1077 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1081 * If dst and source tasks are in the same NUMA group, or not
1082 * in any group then look only at task weights.
1084 if (cur->numa_group == env->p->numa_group) {
1085 imp = taskimp + task_weight(cur, env->src_nid) -
1086 task_weight(cur, env->dst_nid);
1088 * Add some hysteresis to prevent swapping the
1089 * tasks within a group over tiny differences.
1091 if (cur->numa_group)
1095 * Compare the group weights. If a task is all by
1096 * itself (not part of a group), use the task weight
1099 if (env->p->numa_group)
1104 if (cur->numa_group)
1105 imp += group_weight(cur, env->src_nid) -
1106 group_weight(cur, env->dst_nid);
1108 imp += task_weight(cur, env->src_nid) -
1109 task_weight(cur, env->dst_nid);
1113 if (imp < env->best_imp)
1117 /* Is there capacity at our destination? */
1118 if (env->src_stats.has_capacity &&
1119 !env->dst_stats.has_capacity)
1125 /* Balance doesn't matter much if we're running a task per cpu */
1126 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1130 * In the overloaded case, try and keep the load balanced.
1133 dst_load = env->dst_stats.load;
1134 src_load = env->src_stats.load;
1136 /* XXX missing power terms */
1137 load = task_h_load(env->p);
1142 load = task_h_load(cur);
1147 /* make src_load the smaller */
1148 if (dst_load < src_load)
1149 swap(dst_load, src_load);
1151 if (src_load * env->imbalance_pct < dst_load * 100)
1155 task_numa_assign(env, cur, imp);
1160 static void task_numa_find_cpu(struct task_numa_env *env,
1161 long taskimp, long groupimp)
1165 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1166 /* Skip this CPU if the source task cannot migrate */
1167 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1171 task_numa_compare(env, taskimp, groupimp);
1175 static int task_numa_migrate(struct task_struct *p)
1177 struct task_numa_env env = {
1180 .src_cpu = task_cpu(p),
1181 .src_nid = task_node(p),
1183 .imbalance_pct = 112,
1189 struct sched_domain *sd;
1190 unsigned long taskweight, groupweight;
1192 long taskimp, groupimp;
1195 * Pick the lowest SD_NUMA domain, as that would have the smallest
1196 * imbalance and would be the first to start moving tasks about.
1198 * And we want to avoid any moving of tasks about, as that would create
1199 * random movement of tasks -- counter the numa conditions we're trying
1203 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1205 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1209 * Cpusets can break the scheduler domain tree into smaller
1210 * balance domains, some of which do not cross NUMA boundaries.
1211 * Tasks that are "trapped" in such domains cannot be migrated
1212 * elsewhere, so there is no point in (re)trying.
1214 if (unlikely(!sd)) {
1215 p->numa_preferred_nid = cpu_to_node(task_cpu(p));
1219 taskweight = task_weight(p, env.src_nid);
1220 groupweight = group_weight(p, env.src_nid);
1221 update_numa_stats(&env.src_stats, env.src_nid);
1222 env.dst_nid = p->numa_preferred_nid;
1223 taskimp = task_weight(p, env.dst_nid) - taskweight;
1224 groupimp = group_weight(p, env.dst_nid) - groupweight;
1225 update_numa_stats(&env.dst_stats, env.dst_nid);
1227 /* If the preferred nid has capacity, try to use it. */
1228 if (env.dst_stats.has_capacity)
1229 task_numa_find_cpu(&env, taskimp, groupimp);
1231 /* No space available on the preferred nid. Look elsewhere. */
1232 if (env.best_cpu == -1) {
1233 for_each_online_node(nid) {
1234 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1237 /* Only consider nodes where both task and groups benefit */
1238 taskimp = task_weight(p, nid) - taskweight;
1239 groupimp = group_weight(p, nid) - groupweight;
1240 if (taskimp < 0 && groupimp < 0)
1244 update_numa_stats(&env.dst_stats, env.dst_nid);
1245 task_numa_find_cpu(&env, taskimp, groupimp);
1249 /* No better CPU than the current one was found. */
1250 if (env.best_cpu == -1)
1253 sched_setnuma(p, env.dst_nid);
1256 * Reset the scan period if the task is being rescheduled on an
1257 * alternative node to recheck if the tasks is now properly placed.
1259 p->numa_scan_period = task_scan_min(p);
1261 if (env.best_task == NULL) {
1262 int ret = migrate_task_to(p, env.best_cpu);
1266 ret = migrate_swap(p, env.best_task);
1267 put_task_struct(env.best_task);
1271 /* Attempt to migrate a task to a CPU on the preferred node. */
1272 static void numa_migrate_preferred(struct task_struct *p)
1274 /* This task has no NUMA fault statistics yet */
1275 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1278 /* Periodically retry migrating the task to the preferred node */
1279 p->numa_migrate_retry = jiffies + HZ;
1281 /* Success if task is already running on preferred CPU */
1282 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid)
1285 /* Otherwise, try migrate to a CPU on the preferred node */
1286 task_numa_migrate(p);
1290 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1291 * increments. The more local the fault statistics are, the higher the scan
1292 * period will be for the next scan window. If local/remote ratio is below
1293 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1294 * scan period will decrease
1296 #define NUMA_PERIOD_SLOTS 10
1297 #define NUMA_PERIOD_THRESHOLD 3
1300 * Increase the scan period (slow down scanning) if the majority of
1301 * our memory is already on our local node, or if the majority of
1302 * the page accesses are shared with other processes.
1303 * Otherwise, decrease the scan period.
1305 static void update_task_scan_period(struct task_struct *p,
1306 unsigned long shared, unsigned long private)
1308 unsigned int period_slot;
1312 unsigned long remote = p->numa_faults_locality[0];
1313 unsigned long local = p->numa_faults_locality[1];
1316 * If there were no record hinting faults then either the task is
1317 * completely idle or all activity is areas that are not of interest
1318 * to automatic numa balancing. Scan slower
1320 if (local + shared == 0) {
1321 p->numa_scan_period = min(p->numa_scan_period_max,
1322 p->numa_scan_period << 1);
1324 p->mm->numa_next_scan = jiffies +
1325 msecs_to_jiffies(p->numa_scan_period);
1331 * Prepare to scale scan period relative to the current period.
1332 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1333 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1334 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1336 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1337 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1338 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1339 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1342 diff = slot * period_slot;
1344 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1347 * Scale scan rate increases based on sharing. There is an
1348 * inverse relationship between the degree of sharing and
1349 * the adjustment made to the scanning period. Broadly
1350 * speaking the intent is that there is little point
1351 * scanning faster if shared accesses dominate as it may
1352 * simply bounce migrations uselessly
1354 period_slot = DIV_ROUND_UP(diff, NUMA_PERIOD_SLOTS);
1355 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1356 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1359 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1360 task_scan_min(p), task_scan_max(p));
1361 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1364 static void task_numa_placement(struct task_struct *p)
1366 int seq, nid, max_nid = -1, max_group_nid = -1;
1367 unsigned long max_faults = 0, max_group_faults = 0;
1368 unsigned long fault_types[2] = { 0, 0 };
1369 spinlock_t *group_lock = NULL;
1371 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1372 if (p->numa_scan_seq == seq)
1374 p->numa_scan_seq = seq;
1375 p->numa_scan_period_max = task_scan_max(p);
1377 /* If the task is part of a group prevent parallel updates to group stats */
1378 if (p->numa_group) {
1379 group_lock = &p->numa_group->lock;
1380 spin_lock(group_lock);
1383 /* Find the node with the highest number of faults */
1384 for_each_online_node(nid) {
1385 unsigned long faults = 0, group_faults = 0;
1388 for (priv = 0; priv < 2; priv++) {
1391 i = task_faults_idx(nid, priv);
1392 diff = -p->numa_faults[i];
1394 /* Decay existing window, copy faults since last scan */
1395 p->numa_faults[i] >>= 1;
1396 p->numa_faults[i] += p->numa_faults_buffer[i];
1397 fault_types[priv] += p->numa_faults_buffer[i];
1398 p->numa_faults_buffer[i] = 0;
1400 faults += p->numa_faults[i];
1401 diff += p->numa_faults[i];
1402 p->total_numa_faults += diff;
1403 if (p->numa_group) {
1404 /* safe because we can only change our own group */
1405 p->numa_group->faults[i] += diff;
1406 p->numa_group->total_faults += diff;
1407 group_faults += p->numa_group->faults[i];
1411 if (faults > max_faults) {
1412 max_faults = faults;
1416 if (group_faults > max_group_faults) {
1417 max_group_faults = group_faults;
1418 max_group_nid = nid;
1422 update_task_scan_period(p, fault_types[0], fault_types[1]);
1424 if (p->numa_group) {
1426 * If the preferred task and group nids are different,
1427 * iterate over the nodes again to find the best place.
1429 if (max_nid != max_group_nid) {
1430 unsigned long weight, max_weight = 0;
1432 for_each_online_node(nid) {
1433 weight = task_weight(p, nid) + group_weight(p, nid);
1434 if (weight > max_weight) {
1435 max_weight = weight;
1441 spin_unlock(group_lock);
1444 /* Preferred node as the node with the most faults */
1445 if (max_faults && max_nid != p->numa_preferred_nid) {
1446 /* Update the preferred nid and migrate task if possible */
1447 sched_setnuma(p, max_nid);
1448 numa_migrate_preferred(p);
1452 static inline int get_numa_group(struct numa_group *grp)
1454 return atomic_inc_not_zero(&grp->refcount);
1457 static inline void put_numa_group(struct numa_group *grp)
1459 if (atomic_dec_and_test(&grp->refcount))
1460 kfree_rcu(grp, rcu);
1463 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1466 struct numa_group *grp, *my_grp;
1467 struct task_struct *tsk;
1469 int cpu = cpupid_to_cpu(cpupid);
1472 if (unlikely(!p->numa_group)) {
1473 unsigned int size = sizeof(struct numa_group) +
1474 2*nr_node_ids*sizeof(unsigned long);
1476 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1480 atomic_set(&grp->refcount, 1);
1481 spin_lock_init(&grp->lock);
1482 INIT_LIST_HEAD(&grp->task_list);
1485 for (i = 0; i < 2*nr_node_ids; i++)
1486 grp->faults[i] = p->numa_faults[i];
1488 grp->total_faults = p->total_numa_faults;
1490 list_add(&p->numa_entry, &grp->task_list);
1492 rcu_assign_pointer(p->numa_group, grp);
1496 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1498 if (!cpupid_match_pid(tsk, cpupid))
1501 grp = rcu_dereference(tsk->numa_group);
1505 my_grp = p->numa_group;
1510 * Only join the other group if its bigger; if we're the bigger group,
1511 * the other task will join us.
1513 if (my_grp->nr_tasks > grp->nr_tasks)
1517 * Tie-break on the grp address.
1519 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1522 /* Always join threads in the same process. */
1523 if (tsk->mm == current->mm)
1526 /* Simple filter to avoid false positives due to PID collisions */
1527 if (flags & TNF_SHARED)
1530 /* Update priv based on whether false sharing was detected */
1533 if (join && !get_numa_group(grp))
1541 double_lock(&my_grp->lock, &grp->lock);
1543 for (i = 0; i < 2*nr_node_ids; i++) {
1544 my_grp->faults[i] -= p->numa_faults[i];
1545 grp->faults[i] += p->numa_faults[i];
1547 my_grp->total_faults -= p->total_numa_faults;
1548 grp->total_faults += p->total_numa_faults;
1550 list_move(&p->numa_entry, &grp->task_list);
1554 spin_unlock(&my_grp->lock);
1555 spin_unlock(&grp->lock);
1557 rcu_assign_pointer(p->numa_group, grp);
1559 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 spin_lock(&grp->lock);
1575 for (i = 0; i < 2*nr_node_ids; i++)
1576 grp->faults[i] -= p->numa_faults[i];
1577 grp->total_faults -= p->total_numa_faults;
1579 list_del(&p->numa_entry);
1581 spin_unlock(&grp->lock);
1582 rcu_assign_pointer(p->numa_group, NULL);
1583 put_numa_group(grp);
1586 p->numa_faults = NULL;
1587 p->numa_faults_buffer = NULL;
1592 * Got a PROT_NONE fault for a page on @node.
1594 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1596 struct task_struct *p = current;
1597 bool migrated = flags & TNF_MIGRATED;
1600 if (!numabalancing_enabled)
1603 /* for example, ksmd faulting in a user's mm */
1607 /* Do not worry about placement if exiting */
1608 if (p->state == TASK_DEAD)
1611 /* Allocate buffer to track faults on a per-node basis */
1612 if (unlikely(!p->numa_faults)) {
1613 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1615 /* numa_faults and numa_faults_buffer share the allocation */
1616 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1617 if (!p->numa_faults)
1620 BUG_ON(p->numa_faults_buffer);
1621 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1622 p->total_numa_faults = 0;
1623 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1627 * First accesses are treated as private, otherwise consider accesses
1628 * to be private if the accessing pid has not changed
1630 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1633 priv = cpupid_match_pid(p, last_cpupid);
1634 if (!priv && !(flags & TNF_NO_GROUP))
1635 task_numa_group(p, last_cpupid, flags, &priv);
1638 task_numa_placement(p);
1641 * Retry task to preferred node migration periodically, in case it
1642 * case it previously failed, or the scheduler moved us.
1644 if (time_after(jiffies, p->numa_migrate_retry))
1645 numa_migrate_preferred(p);
1648 p->numa_pages_migrated += pages;
1650 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1651 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1654 static void reset_ptenuma_scan(struct task_struct *p)
1656 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1657 p->mm->numa_scan_offset = 0;
1661 * The expensive part of numa migration is done from task_work context.
1662 * Triggered from task_tick_numa().
1664 void task_numa_work(struct callback_head *work)
1666 unsigned long migrate, next_scan, now = jiffies;
1667 struct task_struct *p = current;
1668 struct mm_struct *mm = p->mm;
1669 struct vm_area_struct *vma;
1670 unsigned long start, end;
1671 unsigned long nr_pte_updates = 0;
1674 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1676 work->next = work; /* protect against double add */
1678 * Who cares about NUMA placement when they're dying.
1680 * NOTE: make sure not to dereference p->mm before this check,
1681 * exit_task_work() happens _after_ exit_mm() so we could be called
1682 * without p->mm even though we still had it when we enqueued this
1685 if (p->flags & PF_EXITING)
1688 if (!mm->numa_next_scan) {
1689 mm->numa_next_scan = now +
1690 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1694 * Enforce maximal scan/migration frequency..
1696 migrate = mm->numa_next_scan;
1697 if (time_before(now, migrate))
1700 if (p->numa_scan_period == 0) {
1701 p->numa_scan_period_max = task_scan_max(p);
1702 p->numa_scan_period = task_scan_min(p);
1705 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1706 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1710 * Delay this task enough that another task of this mm will likely win
1711 * the next time around.
1713 p->node_stamp += 2 * TICK_NSEC;
1715 start = mm->numa_scan_offset;
1716 pages = sysctl_numa_balancing_scan_size;
1717 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1721 down_read(&mm->mmap_sem);
1722 vma = find_vma(mm, start);
1724 reset_ptenuma_scan(p);
1728 for (; vma; vma = vma->vm_next) {
1729 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1733 * Shared library pages mapped by multiple processes are not
1734 * migrated as it is expected they are cache replicated. Avoid
1735 * hinting faults in read-only file-backed mappings or the vdso
1736 * as migrating the pages will be of marginal benefit.
1739 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1743 start = max(start, vma->vm_start);
1744 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1745 end = min(end, vma->vm_end);
1746 nr_pte_updates += change_prot_numa(vma, start, end);
1749 * Scan sysctl_numa_balancing_scan_size but ensure that
1750 * at least one PTE is updated so that unused virtual
1751 * address space is quickly skipped.
1754 pages -= (end - start) >> PAGE_SHIFT;
1759 } while (end != vma->vm_end);
1764 * It is possible to reach the end of the VMA list but the last few
1765 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1766 * would find the !migratable VMA on the next scan but not reset the
1767 * scanner to the start so check it now.
1770 mm->numa_scan_offset = start;
1772 reset_ptenuma_scan(p);
1773 up_read(&mm->mmap_sem);
1777 * Drive the periodic memory faults..
1779 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1781 struct callback_head *work = &curr->numa_work;
1785 * We don't care about NUMA placement if we don't have memory.
1787 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1791 * Using runtime rather than walltime has the dual advantage that
1792 * we (mostly) drive the selection from busy threads and that the
1793 * task needs to have done some actual work before we bother with
1796 now = curr->se.sum_exec_runtime;
1797 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1799 if (now - curr->node_stamp > period) {
1800 if (!curr->node_stamp)
1801 curr->numa_scan_period = task_scan_min(curr);
1802 curr->node_stamp += period;
1804 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1805 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1806 task_work_add(curr, work, true);
1811 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1815 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1819 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1822 #endif /* CONFIG_NUMA_BALANCING */
1825 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1827 update_load_add(&cfs_rq->load, se->load.weight);
1828 if (!parent_entity(se))
1829 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1831 if (entity_is_task(se)) {
1832 struct rq *rq = rq_of(cfs_rq);
1834 account_numa_enqueue(rq, task_of(se));
1835 list_add(&se->group_node, &rq->cfs_tasks);
1838 cfs_rq->nr_running++;
1842 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1844 update_load_sub(&cfs_rq->load, se->load.weight);
1845 if (!parent_entity(se))
1846 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1847 if (entity_is_task(se)) {
1848 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1849 list_del_init(&se->group_node);
1851 cfs_rq->nr_running--;
1854 #ifdef CONFIG_FAIR_GROUP_SCHED
1856 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1861 * Use this CPU's actual weight instead of the last load_contribution
1862 * to gain a more accurate current total weight. See
1863 * update_cfs_rq_load_contribution().
1865 tg_weight = atomic_long_read(&tg->load_avg);
1866 tg_weight -= cfs_rq->tg_load_contrib;
1867 tg_weight += cfs_rq->load.weight;
1872 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1874 long tg_weight, load, shares;
1876 tg_weight = calc_tg_weight(tg, cfs_rq);
1877 load = cfs_rq->load.weight;
1879 shares = (tg->shares * load);
1881 shares /= tg_weight;
1883 if (shares < MIN_SHARES)
1884 shares = MIN_SHARES;
1885 if (shares > tg->shares)
1886 shares = tg->shares;
1890 # else /* CONFIG_SMP */
1891 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1895 # endif /* CONFIG_SMP */
1896 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1897 unsigned long weight)
1900 /* commit outstanding execution time */
1901 if (cfs_rq->curr == se)
1902 update_curr(cfs_rq);
1903 account_entity_dequeue(cfs_rq, se);
1906 update_load_set(&se->load, weight);
1909 account_entity_enqueue(cfs_rq, se);
1912 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1914 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1916 struct task_group *tg;
1917 struct sched_entity *se;
1921 se = tg->se[cpu_of(rq_of(cfs_rq))];
1922 if (!se || throttled_hierarchy(cfs_rq))
1925 if (likely(se->load.weight == tg->shares))
1928 shares = calc_cfs_shares(cfs_rq, tg);
1930 reweight_entity(cfs_rq_of(se), se, shares);
1932 #else /* CONFIG_FAIR_GROUP_SCHED */
1933 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1936 #endif /* CONFIG_FAIR_GROUP_SCHED */
1940 * We choose a half-life close to 1 scheduling period.
1941 * Note: The tables below are dependent on this value.
1943 #define LOAD_AVG_PERIOD 32
1944 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1945 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1947 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1948 static const u32 runnable_avg_yN_inv[] = {
1949 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1950 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1951 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1952 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1953 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1954 0x85aac367, 0x82cd8698,
1958 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1959 * over-estimates when re-combining.
1961 static const u32 runnable_avg_yN_sum[] = {
1962 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1963 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1964 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1969 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1971 static __always_inline u64 decay_load(u64 val, u64 n)
1973 unsigned int local_n;
1977 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1980 /* after bounds checking we can collapse to 32-bit */
1984 * As y^PERIOD = 1/2, we can combine
1985 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1986 * With a look-up table which covers k^n (n<PERIOD)
1988 * To achieve constant time decay_load.
1990 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1991 val >>= local_n / LOAD_AVG_PERIOD;
1992 local_n %= LOAD_AVG_PERIOD;
1995 val *= runnable_avg_yN_inv[local_n];
1996 /* We don't use SRR here since we always want to round down. */
2001 * For updates fully spanning n periods, the contribution to runnable
2002 * average will be: \Sum 1024*y^n
2004 * We can compute this reasonably efficiently by combining:
2005 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2007 static u32 __compute_runnable_contrib(u64 n)
2011 if (likely(n <= LOAD_AVG_PERIOD))
2012 return runnable_avg_yN_sum[n];
2013 else if (unlikely(n >= LOAD_AVG_MAX_N))
2014 return LOAD_AVG_MAX;
2016 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2018 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2019 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2021 n -= LOAD_AVG_PERIOD;
2022 } while (n > LOAD_AVG_PERIOD);
2024 contrib = decay_load(contrib, n);
2025 return contrib + runnable_avg_yN_sum[n];
2029 * We can represent the historical contribution to runnable average as the
2030 * coefficients of a geometric series. To do this we sub-divide our runnable
2031 * history into segments of approximately 1ms (1024us); label the segment that
2032 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2034 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2036 * (now) (~1ms ago) (~2ms ago)
2038 * Let u_i denote the fraction of p_i that the entity was runnable.
2040 * We then designate the fractions u_i as our co-efficients, yielding the
2041 * following representation of historical load:
2042 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2044 * We choose y based on the with of a reasonably scheduling period, fixing:
2047 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2048 * approximately half as much as the contribution to load within the last ms
2051 * When a period "rolls over" and we have new u_0`, multiplying the previous
2052 * sum again by y is sufficient to update:
2053 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2054 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2056 static __always_inline int __update_entity_runnable_avg(u64 now,
2057 struct sched_avg *sa,
2061 u32 runnable_contrib;
2062 int delta_w, decayed = 0;
2064 delta = now - sa->last_runnable_update;
2066 * This should only happen when time goes backwards, which it
2067 * unfortunately does during sched clock init when we swap over to TSC.
2069 if ((s64)delta < 0) {
2070 sa->last_runnable_update = now;
2075 * Use 1024ns as the unit of measurement since it's a reasonable
2076 * approximation of 1us and fast to compute.
2081 sa->last_runnable_update = now;
2083 /* delta_w is the amount already accumulated against our next period */
2084 delta_w = sa->runnable_avg_period % 1024;
2085 if (delta + delta_w >= 1024) {
2086 /* period roll-over */
2090 * Now that we know we're crossing a period boundary, figure
2091 * out how much from delta we need to complete the current
2092 * period and accrue it.
2094 delta_w = 1024 - delta_w;
2096 sa->runnable_avg_sum += delta_w;
2097 sa->runnable_avg_period += delta_w;
2101 /* Figure out how many additional periods this update spans */
2102 periods = delta / 1024;
2105 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2107 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2110 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2111 runnable_contrib = __compute_runnable_contrib(periods);
2113 sa->runnable_avg_sum += runnable_contrib;
2114 sa->runnable_avg_period += runnable_contrib;
2117 /* Remainder of delta accrued against u_0` */
2119 sa->runnable_avg_sum += delta;
2120 sa->runnable_avg_period += delta;
2125 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2126 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2128 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2129 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2131 decays -= se->avg.decay_count;
2135 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2136 se->avg.decay_count = 0;
2141 #ifdef CONFIG_FAIR_GROUP_SCHED
2142 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2145 struct task_group *tg = cfs_rq->tg;
2148 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2149 tg_contrib -= cfs_rq->tg_load_contrib;
2151 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2152 atomic_long_add(tg_contrib, &tg->load_avg);
2153 cfs_rq->tg_load_contrib += tg_contrib;
2158 * Aggregate cfs_rq runnable averages into an equivalent task_group
2159 * representation for computing load contributions.
2161 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2162 struct cfs_rq *cfs_rq)
2164 struct task_group *tg = cfs_rq->tg;
2167 /* The fraction of a cpu used by this cfs_rq */
2168 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2169 sa->runnable_avg_period + 1);
2170 contrib -= cfs_rq->tg_runnable_contrib;
2172 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2173 atomic_add(contrib, &tg->runnable_avg);
2174 cfs_rq->tg_runnable_contrib += contrib;
2178 static inline void __update_group_entity_contrib(struct sched_entity *se)
2180 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2181 struct task_group *tg = cfs_rq->tg;
2186 contrib = cfs_rq->tg_load_contrib * tg->shares;
2187 se->avg.load_avg_contrib = div_u64(contrib,
2188 atomic_long_read(&tg->load_avg) + 1);
2191 * For group entities we need to compute a correction term in the case
2192 * that they are consuming <1 cpu so that we would contribute the same
2193 * load as a task of equal weight.
2195 * Explicitly co-ordinating this measurement would be expensive, but
2196 * fortunately the sum of each cpus contribution forms a usable
2197 * lower-bound on the true value.
2199 * Consider the aggregate of 2 contributions. Either they are disjoint
2200 * (and the sum represents true value) or they are disjoint and we are
2201 * understating by the aggregate of their overlap.
2203 * Extending this to N cpus, for a given overlap, the maximum amount we
2204 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2205 * cpus that overlap for this interval and w_i is the interval width.
2207 * On a small machine; the first term is well-bounded which bounds the
2208 * total error since w_i is a subset of the period. Whereas on a
2209 * larger machine, while this first term can be larger, if w_i is the
2210 * of consequential size guaranteed to see n_i*w_i quickly converge to
2211 * our upper bound of 1-cpu.
2213 runnable_avg = atomic_read(&tg->runnable_avg);
2214 if (runnable_avg < NICE_0_LOAD) {
2215 se->avg.load_avg_contrib *= runnable_avg;
2216 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2220 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2221 int force_update) {}
2222 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2223 struct cfs_rq *cfs_rq) {}
2224 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2227 static inline void __update_task_entity_contrib(struct sched_entity *se)
2231 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2232 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2233 contrib /= (se->avg.runnable_avg_period + 1);
2234 se->avg.load_avg_contrib = scale_load(contrib);
2237 /* Compute the current contribution to load_avg by se, return any delta */
2238 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2240 long old_contrib = se->avg.load_avg_contrib;
2242 if (entity_is_task(se)) {
2243 __update_task_entity_contrib(se);
2245 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2246 __update_group_entity_contrib(se);
2249 return se->avg.load_avg_contrib - old_contrib;
2252 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2255 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2256 cfs_rq->blocked_load_avg -= load_contrib;
2258 cfs_rq->blocked_load_avg = 0;
2261 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2263 /* Update a sched_entity's runnable average */
2264 static inline void update_entity_load_avg(struct sched_entity *se,
2267 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2272 * For a group entity we need to use their owned cfs_rq_clock_task() in
2273 * case they are the parent of a throttled hierarchy.
2275 if (entity_is_task(se))
2276 now = cfs_rq_clock_task(cfs_rq);
2278 now = cfs_rq_clock_task(group_cfs_rq(se));
2280 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2283 contrib_delta = __update_entity_load_avg_contrib(se);
2289 cfs_rq->runnable_load_avg += contrib_delta;
2291 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2295 * Decay the load contributed by all blocked children and account this so that
2296 * their contribution may appropriately discounted when they wake up.
2298 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2300 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2303 decays = now - cfs_rq->last_decay;
2304 if (!decays && !force_update)
2307 if (atomic_long_read(&cfs_rq->removed_load)) {
2308 unsigned long removed_load;
2309 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2310 subtract_blocked_load_contrib(cfs_rq, removed_load);
2314 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2316 atomic64_add(decays, &cfs_rq->decay_counter);
2317 cfs_rq->last_decay = now;
2320 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2323 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2325 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2326 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2329 /* Add the load generated by se into cfs_rq's child load-average */
2330 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2331 struct sched_entity *se,
2335 * We track migrations using entity decay_count <= 0, on a wake-up
2336 * migration we use a negative decay count to track the remote decays
2337 * accumulated while sleeping.
2339 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2340 * are seen by enqueue_entity_load_avg() as a migration with an already
2341 * constructed load_avg_contrib.
2343 if (unlikely(se->avg.decay_count <= 0)) {
2344 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2345 if (se->avg.decay_count) {
2347 * In a wake-up migration we have to approximate the
2348 * time sleeping. This is because we can't synchronize
2349 * clock_task between the two cpus, and it is not
2350 * guaranteed to be read-safe. Instead, we can
2351 * approximate this using our carried decays, which are
2352 * explicitly atomically readable.
2354 se->avg.last_runnable_update -= (-se->avg.decay_count)
2356 update_entity_load_avg(se, 0);
2357 /* Indicate that we're now synchronized and on-rq */
2358 se->avg.decay_count = 0;
2363 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2364 * would have made count negative); we must be careful to avoid
2365 * double-accounting blocked time after synchronizing decays.
2367 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2371 /* migrated tasks did not contribute to our blocked load */
2373 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2374 update_entity_load_avg(se, 0);
2377 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2378 /* we force update consideration on load-balancer moves */
2379 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2383 * Remove se's load from this cfs_rq child load-average, if the entity is
2384 * transitioning to a blocked state we track its projected decay using
2387 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2388 struct sched_entity *se,
2391 update_entity_load_avg(se, 1);
2392 /* we force update consideration on load-balancer moves */
2393 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2395 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2397 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2398 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2399 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2403 * Update the rq's load with the elapsed running time before entering
2404 * idle. if the last scheduled task is not a CFS task, idle_enter will
2405 * be the only way to update the runnable statistic.
2407 void idle_enter_fair(struct rq *this_rq)
2409 update_rq_runnable_avg(this_rq, 1);
2413 * Update the rq's load with the elapsed idle time before a task is
2414 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2415 * be the only way to update the runnable statistic.
2417 void idle_exit_fair(struct rq *this_rq)
2419 update_rq_runnable_avg(this_rq, 0);
2423 static inline void update_entity_load_avg(struct sched_entity *se,
2424 int update_cfs_rq) {}
2425 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2426 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2427 struct sched_entity *se,
2429 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2430 struct sched_entity *se,
2432 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2433 int force_update) {}
2436 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2438 #ifdef CONFIG_SCHEDSTATS
2439 struct task_struct *tsk = NULL;
2441 if (entity_is_task(se))
2444 if (se->statistics.sleep_start) {
2445 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2450 if (unlikely(delta > se->statistics.sleep_max))
2451 se->statistics.sleep_max = delta;
2453 se->statistics.sleep_start = 0;
2454 se->statistics.sum_sleep_runtime += delta;
2457 account_scheduler_latency(tsk, delta >> 10, 1);
2458 trace_sched_stat_sleep(tsk, delta);
2461 if (se->statistics.block_start) {
2462 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2467 if (unlikely(delta > se->statistics.block_max))
2468 se->statistics.block_max = delta;
2470 se->statistics.block_start = 0;
2471 se->statistics.sum_sleep_runtime += delta;
2474 if (tsk->in_iowait) {
2475 se->statistics.iowait_sum += delta;
2476 se->statistics.iowait_count++;
2477 trace_sched_stat_iowait(tsk, delta);
2480 trace_sched_stat_blocked(tsk, delta);
2483 * Blocking time is in units of nanosecs, so shift by
2484 * 20 to get a milliseconds-range estimation of the
2485 * amount of time that the task spent sleeping:
2487 if (unlikely(prof_on == SLEEP_PROFILING)) {
2488 profile_hits(SLEEP_PROFILING,
2489 (void *)get_wchan(tsk),
2492 account_scheduler_latency(tsk, delta >> 10, 0);
2498 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2500 #ifdef CONFIG_SCHED_DEBUG
2501 s64 d = se->vruntime - cfs_rq->min_vruntime;
2506 if (d > 3*sysctl_sched_latency)
2507 schedstat_inc(cfs_rq, nr_spread_over);
2512 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2514 u64 vruntime = cfs_rq->min_vruntime;
2517 * The 'current' period is already promised to the current tasks,
2518 * however the extra weight of the new task will slow them down a
2519 * little, place the new task so that it fits in the slot that
2520 * stays open at the end.
2522 if (initial && sched_feat(START_DEBIT))
2523 vruntime += sched_vslice(cfs_rq, se);
2525 /* sleeps up to a single latency don't count. */
2527 unsigned long thresh = sysctl_sched_latency;
2530 * Halve their sleep time's effect, to allow
2531 * for a gentler effect of sleepers:
2533 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2539 /* ensure we never gain time by being placed backwards. */
2540 se->vruntime = max_vruntime(se->vruntime, vruntime);
2543 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2546 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2549 * Update the normalized vruntime before updating min_vruntime
2550 * through calling update_curr().
2552 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2553 se->vruntime += cfs_rq->min_vruntime;
2556 * Update run-time statistics of the 'current'.
2558 update_curr(cfs_rq);
2559 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2560 account_entity_enqueue(cfs_rq, se);
2561 update_cfs_shares(cfs_rq);
2563 if (flags & ENQUEUE_WAKEUP) {
2564 place_entity(cfs_rq, se, 0);
2565 enqueue_sleeper(cfs_rq, se);
2568 update_stats_enqueue(cfs_rq, se);
2569 check_spread(cfs_rq, se);
2570 if (se != cfs_rq->curr)
2571 __enqueue_entity(cfs_rq, se);
2574 if (cfs_rq->nr_running == 1) {
2575 list_add_leaf_cfs_rq(cfs_rq);
2576 check_enqueue_throttle(cfs_rq);
2580 static void __clear_buddies_last(struct sched_entity *se)
2582 for_each_sched_entity(se) {
2583 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2584 if (cfs_rq->last == se)
2585 cfs_rq->last = NULL;
2591 static void __clear_buddies_next(struct sched_entity *se)
2593 for_each_sched_entity(se) {
2594 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2595 if (cfs_rq->next == se)
2596 cfs_rq->next = NULL;
2602 static void __clear_buddies_skip(struct sched_entity *se)
2604 for_each_sched_entity(se) {
2605 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2606 if (cfs_rq->skip == se)
2607 cfs_rq->skip = NULL;
2613 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2615 if (cfs_rq->last == se)
2616 __clear_buddies_last(se);
2618 if (cfs_rq->next == se)
2619 __clear_buddies_next(se);
2621 if (cfs_rq->skip == se)
2622 __clear_buddies_skip(se);
2625 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2628 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2631 * Update run-time statistics of the 'current'.
2633 update_curr(cfs_rq);
2634 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2636 update_stats_dequeue(cfs_rq, se);
2637 if (flags & DEQUEUE_SLEEP) {
2638 #ifdef CONFIG_SCHEDSTATS
2639 if (entity_is_task(se)) {
2640 struct task_struct *tsk = task_of(se);
2642 if (tsk->state & TASK_INTERRUPTIBLE)
2643 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2644 if (tsk->state & TASK_UNINTERRUPTIBLE)
2645 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2650 clear_buddies(cfs_rq, se);
2652 if (se != cfs_rq->curr)
2653 __dequeue_entity(cfs_rq, se);
2655 account_entity_dequeue(cfs_rq, se);
2658 * Normalize the entity after updating the min_vruntime because the
2659 * update can refer to the ->curr item and we need to reflect this
2660 * movement in our normalized position.
2662 if (!(flags & DEQUEUE_SLEEP))
2663 se->vruntime -= cfs_rq->min_vruntime;
2665 /* return excess runtime on last dequeue */
2666 return_cfs_rq_runtime(cfs_rq);
2668 update_min_vruntime(cfs_rq);
2669 update_cfs_shares(cfs_rq);
2673 * Preempt the current task with a newly woken task if needed:
2676 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2678 unsigned long ideal_runtime, delta_exec;
2679 struct sched_entity *se;
2682 ideal_runtime = sched_slice(cfs_rq, curr);
2683 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2684 if (delta_exec > ideal_runtime) {
2685 resched_task(rq_of(cfs_rq)->curr);
2687 * The current task ran long enough, ensure it doesn't get
2688 * re-elected due to buddy favours.
2690 clear_buddies(cfs_rq, curr);
2695 * Ensure that a task that missed wakeup preemption by a
2696 * narrow margin doesn't have to wait for a full slice.
2697 * This also mitigates buddy induced latencies under load.
2699 if (delta_exec < sysctl_sched_min_granularity)
2702 se = __pick_first_entity(cfs_rq);
2703 delta = curr->vruntime - se->vruntime;
2708 if (delta > ideal_runtime)
2709 resched_task(rq_of(cfs_rq)->curr);
2713 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2715 /* 'current' is not kept within the tree. */
2718 * Any task has to be enqueued before it get to execute on
2719 * a CPU. So account for the time it spent waiting on the
2722 update_stats_wait_end(cfs_rq, se);
2723 __dequeue_entity(cfs_rq, se);
2726 update_stats_curr_start(cfs_rq, se);
2728 #ifdef CONFIG_SCHEDSTATS
2730 * Track our maximum slice length, if the CPU's load is at
2731 * least twice that of our own weight (i.e. dont track it
2732 * when there are only lesser-weight tasks around):
2734 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2735 se->statistics.slice_max = max(se->statistics.slice_max,
2736 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2739 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2743 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2746 * Pick the next process, keeping these things in mind, in this order:
2747 * 1) keep things fair between processes/task groups
2748 * 2) pick the "next" process, since someone really wants that to run
2749 * 3) pick the "last" process, for cache locality
2750 * 4) do not run the "skip" process, if something else is available
2752 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2754 struct sched_entity *se = __pick_first_entity(cfs_rq);
2755 struct sched_entity *left = se;
2758 * Avoid running the skip buddy, if running something else can
2759 * be done without getting too unfair.
2761 if (cfs_rq->skip == se) {
2762 struct sched_entity *second = __pick_next_entity(se);
2763 if (second && wakeup_preempt_entity(second, left) < 1)
2768 * Prefer last buddy, try to return the CPU to a preempted task.
2770 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2774 * Someone really wants this to run. If it's not unfair, run it.
2776 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2779 clear_buddies(cfs_rq, se);
2784 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2786 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2789 * If still on the runqueue then deactivate_task()
2790 * was not called and update_curr() has to be done:
2793 update_curr(cfs_rq);
2795 /* throttle cfs_rqs exceeding runtime */
2796 check_cfs_rq_runtime(cfs_rq);
2798 check_spread(cfs_rq, prev);
2800 update_stats_wait_start(cfs_rq, prev);
2801 /* Put 'current' back into the tree. */
2802 __enqueue_entity(cfs_rq, prev);
2803 /* in !on_rq case, update occurred at dequeue */
2804 update_entity_load_avg(prev, 1);
2806 cfs_rq->curr = NULL;
2810 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2813 * Update run-time statistics of the 'current'.
2815 update_curr(cfs_rq);
2818 * Ensure that runnable average is periodically updated.
2820 update_entity_load_avg(curr, 1);
2821 update_cfs_rq_blocked_load(cfs_rq, 1);
2822 update_cfs_shares(cfs_rq);
2824 #ifdef CONFIG_SCHED_HRTICK
2826 * queued ticks are scheduled to match the slice, so don't bother
2827 * validating it and just reschedule.
2830 resched_task(rq_of(cfs_rq)->curr);
2834 * don't let the period tick interfere with the hrtick preemption
2836 if (!sched_feat(DOUBLE_TICK) &&
2837 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2841 if (cfs_rq->nr_running > 1)
2842 check_preempt_tick(cfs_rq, curr);
2846 /**************************************************
2847 * CFS bandwidth control machinery
2850 #ifdef CONFIG_CFS_BANDWIDTH
2852 #ifdef HAVE_JUMP_LABEL
2853 static struct static_key __cfs_bandwidth_used;
2855 static inline bool cfs_bandwidth_used(void)
2857 return static_key_false(&__cfs_bandwidth_used);
2860 void cfs_bandwidth_usage_inc(void)
2862 static_key_slow_inc(&__cfs_bandwidth_used);
2865 void cfs_bandwidth_usage_dec(void)
2867 static_key_slow_dec(&__cfs_bandwidth_used);
2869 #else /* HAVE_JUMP_LABEL */
2870 static bool cfs_bandwidth_used(void)
2875 void cfs_bandwidth_usage_inc(void) {}
2876 void cfs_bandwidth_usage_dec(void) {}
2877 #endif /* HAVE_JUMP_LABEL */
2880 * default period for cfs group bandwidth.
2881 * default: 0.1s, units: nanoseconds
2883 static inline u64 default_cfs_period(void)
2885 return 100000000ULL;
2888 static inline u64 sched_cfs_bandwidth_slice(void)
2890 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2894 * Replenish runtime according to assigned quota and update expiration time.
2895 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2896 * additional synchronization around rq->lock.
2898 * requires cfs_b->lock
2900 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2904 if (cfs_b->quota == RUNTIME_INF)
2907 now = sched_clock_cpu(smp_processor_id());
2908 cfs_b->runtime = cfs_b->quota;
2909 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2912 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2914 return &tg->cfs_bandwidth;
2917 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2918 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2920 if (unlikely(cfs_rq->throttle_count))
2921 return cfs_rq->throttled_clock_task;
2923 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2926 /* returns 0 on failure to allocate runtime */
2927 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2929 struct task_group *tg = cfs_rq->tg;
2930 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2931 u64 amount = 0, min_amount, expires;
2933 /* note: this is a positive sum as runtime_remaining <= 0 */
2934 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2936 raw_spin_lock(&cfs_b->lock);
2937 if (cfs_b->quota == RUNTIME_INF)
2938 amount = min_amount;
2941 * If the bandwidth pool has become inactive, then at least one
2942 * period must have elapsed since the last consumption.
2943 * Refresh the global state and ensure bandwidth timer becomes
2946 if (!cfs_b->timer_active) {
2947 __refill_cfs_bandwidth_runtime(cfs_b);
2948 __start_cfs_bandwidth(cfs_b);
2951 if (cfs_b->runtime > 0) {
2952 amount = min(cfs_b->runtime, min_amount);
2953 cfs_b->runtime -= amount;
2957 expires = cfs_b->runtime_expires;
2958 raw_spin_unlock(&cfs_b->lock);
2960 cfs_rq->runtime_remaining += amount;
2962 * we may have advanced our local expiration to account for allowed
2963 * spread between our sched_clock and the one on which runtime was
2966 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2967 cfs_rq->runtime_expires = expires;
2969 return cfs_rq->runtime_remaining > 0;
2973 * Note: This depends on the synchronization provided by sched_clock and the
2974 * fact that rq->clock snapshots this value.
2976 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2978 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2980 /* if the deadline is ahead of our clock, nothing to do */
2981 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2984 if (cfs_rq->runtime_remaining < 0)
2988 * If the local deadline has passed we have to consider the
2989 * possibility that our sched_clock is 'fast' and the global deadline
2990 * has not truly expired.
2992 * Fortunately we can check determine whether this the case by checking
2993 * whether the global deadline has advanced.
2996 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2997 /* extend local deadline, drift is bounded above by 2 ticks */
2998 cfs_rq->runtime_expires += TICK_NSEC;
3000 /* global deadline is ahead, expiration has passed */
3001 cfs_rq->runtime_remaining = 0;
3005 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3006 unsigned long delta_exec)
3008 /* dock delta_exec before expiring quota (as it could span periods) */
3009 cfs_rq->runtime_remaining -= delta_exec;
3010 expire_cfs_rq_runtime(cfs_rq);
3012 if (likely(cfs_rq->runtime_remaining > 0))
3016 * if we're unable to extend our runtime we resched so that the active
3017 * hierarchy can be throttled
3019 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3020 resched_task(rq_of(cfs_rq)->curr);
3023 static __always_inline
3024 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
3026 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3029 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3032 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3034 return cfs_bandwidth_used() && cfs_rq->throttled;
3037 /* check whether cfs_rq, or any parent, is throttled */
3038 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3040 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3044 * Ensure that neither of the group entities corresponding to src_cpu or
3045 * dest_cpu are members of a throttled hierarchy when performing group
3046 * load-balance operations.
3048 static inline int throttled_lb_pair(struct task_group *tg,
3049 int src_cpu, int dest_cpu)
3051 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3053 src_cfs_rq = tg->cfs_rq[src_cpu];
3054 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3056 return throttled_hierarchy(src_cfs_rq) ||
3057 throttled_hierarchy(dest_cfs_rq);
3060 /* updated child weight may affect parent so we have to do this bottom up */
3061 static int tg_unthrottle_up(struct task_group *tg, void *data)
3063 struct rq *rq = data;
3064 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3066 cfs_rq->throttle_count--;
3068 if (!cfs_rq->throttle_count) {
3069 /* adjust cfs_rq_clock_task() */
3070 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3071 cfs_rq->throttled_clock_task;
3078 static int tg_throttle_down(struct task_group *tg, void *data)
3080 struct rq *rq = data;
3081 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3083 /* group is entering throttled state, stop time */
3084 if (!cfs_rq->throttle_count)
3085 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3086 cfs_rq->throttle_count++;
3091 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3093 struct rq *rq = rq_of(cfs_rq);
3094 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3095 struct sched_entity *se;
3096 long task_delta, dequeue = 1;
3098 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3100 /* freeze hierarchy runnable averages while throttled */
3102 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3105 task_delta = cfs_rq->h_nr_running;
3106 for_each_sched_entity(se) {
3107 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3108 /* throttled entity or throttle-on-deactivate */
3113 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3114 qcfs_rq->h_nr_running -= task_delta;
3116 if (qcfs_rq->load.weight)
3121 rq->nr_running -= task_delta;
3123 cfs_rq->throttled = 1;
3124 cfs_rq->throttled_clock = rq_clock(rq);
3125 raw_spin_lock(&cfs_b->lock);
3126 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3127 if (!cfs_b->timer_active)
3128 __start_cfs_bandwidth(cfs_b);
3129 raw_spin_unlock(&cfs_b->lock);
3132 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3134 struct rq *rq = rq_of(cfs_rq);
3135 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3136 struct sched_entity *se;
3140 se = cfs_rq->tg->se[cpu_of(rq)];
3142 cfs_rq->throttled = 0;
3144 update_rq_clock(rq);
3146 raw_spin_lock(&cfs_b->lock);
3147 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3148 list_del_rcu(&cfs_rq->throttled_list);
3149 raw_spin_unlock(&cfs_b->lock);
3151 /* update hierarchical throttle state */
3152 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3154 if (!cfs_rq->load.weight)
3157 task_delta = cfs_rq->h_nr_running;
3158 for_each_sched_entity(se) {
3162 cfs_rq = cfs_rq_of(se);
3164 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3165 cfs_rq->h_nr_running += task_delta;
3167 if (cfs_rq_throttled(cfs_rq))
3172 rq->nr_running += task_delta;
3174 /* determine whether we need to wake up potentially idle cpu */
3175 if (rq->curr == rq->idle && rq->cfs.nr_running)
3176 resched_task(rq->curr);
3179 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3180 u64 remaining, u64 expires)
3182 struct cfs_rq *cfs_rq;
3183 u64 runtime = remaining;
3186 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3188 struct rq *rq = rq_of(cfs_rq);
3190 raw_spin_lock(&rq->lock);
3191 if (!cfs_rq_throttled(cfs_rq))
3194 runtime = -cfs_rq->runtime_remaining + 1;
3195 if (runtime > remaining)
3196 runtime = remaining;
3197 remaining -= runtime;
3199 cfs_rq->runtime_remaining += runtime;
3200 cfs_rq->runtime_expires = expires;
3202 /* we check whether we're throttled above */
3203 if (cfs_rq->runtime_remaining > 0)
3204 unthrottle_cfs_rq(cfs_rq);
3207 raw_spin_unlock(&rq->lock);
3218 * Responsible for refilling a task_group's bandwidth and unthrottling its
3219 * cfs_rqs as appropriate. If there has been no activity within the last
3220 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3221 * used to track this state.
3223 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3225 u64 runtime, runtime_expires;
3226 int idle = 1, throttled;
3228 raw_spin_lock(&cfs_b->lock);
3229 /* no need to continue the timer with no bandwidth constraint */
3230 if (cfs_b->quota == RUNTIME_INF)
3233 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3234 /* idle depends on !throttled (for the case of a large deficit) */
3235 idle = cfs_b->idle && !throttled;
3236 cfs_b->nr_periods += overrun;
3238 /* if we're going inactive then everything else can be deferred */
3243 * if we have relooped after returning idle once, we need to update our
3244 * status as actually running, so that other cpus doing
3245 * __start_cfs_bandwidth will stop trying to cancel us.
3247 cfs_b->timer_active = 1;
3249 __refill_cfs_bandwidth_runtime(cfs_b);
3252 /* mark as potentially idle for the upcoming period */
3257 /* account preceding periods in which throttling occurred */
3258 cfs_b->nr_throttled += overrun;
3261 * There are throttled entities so we must first use the new bandwidth
3262 * to unthrottle them before making it generally available. This
3263 * ensures that all existing debts will be paid before a new cfs_rq is
3266 runtime = cfs_b->runtime;
3267 runtime_expires = cfs_b->runtime_expires;
3271 * This check is repeated as we are holding onto the new bandwidth
3272 * while we unthrottle. This can potentially race with an unthrottled
3273 * group trying to acquire new bandwidth from the global pool.
3275 while (throttled && runtime > 0) {
3276 raw_spin_unlock(&cfs_b->lock);
3277 /* we can't nest cfs_b->lock while distributing bandwidth */
3278 runtime = distribute_cfs_runtime(cfs_b, runtime,
3280 raw_spin_lock(&cfs_b->lock);
3282 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3285 /* return (any) remaining runtime */
3286 cfs_b->runtime = runtime;
3288 * While we are ensured activity in the period following an
3289 * unthrottle, this also covers the case in which the new bandwidth is
3290 * insufficient to cover the existing bandwidth deficit. (Forcing the
3291 * timer to remain active while there are any throttled entities.)
3296 cfs_b->timer_active = 0;
3297 raw_spin_unlock(&cfs_b->lock);
3302 /* a cfs_rq won't donate quota below this amount */
3303 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3304 /* minimum remaining period time to redistribute slack quota */
3305 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3306 /* how long we wait to gather additional slack before distributing */
3307 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3310 * Are we near the end of the current quota period?
3312 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
3313 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of
3314 * migrate_hrtimers, base is never cleared, so we are fine.
3316 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3318 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3321 /* if the call-back is running a quota refresh is already occurring */
3322 if (hrtimer_callback_running(refresh_timer))
3325 /* is a quota refresh about to occur? */
3326 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3327 if (remaining < min_expire)
3333 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3335 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3337 /* if there's a quota refresh soon don't bother with slack */
3338 if (runtime_refresh_within(cfs_b, min_left))
3341 start_bandwidth_timer(&cfs_b->slack_timer,
3342 ns_to_ktime(cfs_bandwidth_slack_period));
3345 /* we know any runtime found here is valid as update_curr() precedes return */
3346 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3348 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3349 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3351 if (slack_runtime <= 0)
3354 raw_spin_lock(&cfs_b->lock);
3355 if (cfs_b->quota != RUNTIME_INF &&
3356 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3357 cfs_b->runtime += slack_runtime;
3359 /* we are under rq->lock, defer unthrottling using a timer */
3360 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3361 !list_empty(&cfs_b->throttled_cfs_rq))
3362 start_cfs_slack_bandwidth(cfs_b);
3364 raw_spin_unlock(&cfs_b->lock);
3366 /* even if it's not valid for return we don't want to try again */
3367 cfs_rq->runtime_remaining -= slack_runtime;
3370 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3372 if (!cfs_bandwidth_used())
3375 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3378 __return_cfs_rq_runtime(cfs_rq);
3382 * This is done with a timer (instead of inline with bandwidth return) since
3383 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3385 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3387 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3390 /* confirm we're still not at a refresh boundary */
3391 raw_spin_lock(&cfs_b->lock);
3392 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
3393 raw_spin_unlock(&cfs_b->lock);
3397 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3398 runtime = cfs_b->runtime;
3401 expires = cfs_b->runtime_expires;
3402 raw_spin_unlock(&cfs_b->lock);
3407 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3409 raw_spin_lock(&cfs_b->lock);
3410 if (expires == cfs_b->runtime_expires)
3411 cfs_b->runtime = runtime;
3412 raw_spin_unlock(&cfs_b->lock);
3416 * When a group wakes up we want to make sure that its quota is not already
3417 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3418 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3420 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3422 if (!cfs_bandwidth_used())
3425 /* an active group must be handled by the update_curr()->put() path */
3426 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3429 /* ensure the group is not already throttled */
3430 if (cfs_rq_throttled(cfs_rq))
3433 /* update runtime allocation */
3434 account_cfs_rq_runtime(cfs_rq, 0);
3435 if (cfs_rq->runtime_remaining <= 0)
3436 throttle_cfs_rq(cfs_rq);
3439 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3440 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3442 if (!cfs_bandwidth_used())
3445 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3449 * it's possible for a throttled entity to be forced into a running
3450 * state (e.g. set_curr_task), in this case we're finished.
3452 if (cfs_rq_throttled(cfs_rq))
3455 throttle_cfs_rq(cfs_rq);
3458 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3460 struct cfs_bandwidth *cfs_b =
3461 container_of(timer, struct cfs_bandwidth, slack_timer);
3462 do_sched_cfs_slack_timer(cfs_b);
3464 return HRTIMER_NORESTART;
3467 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3469 struct cfs_bandwidth *cfs_b =
3470 container_of(timer, struct cfs_bandwidth, period_timer);
3476 now = hrtimer_cb_get_time(timer);
3477 overrun = hrtimer_forward(timer, now, cfs_b->period);
3482 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3485 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3488 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3490 raw_spin_lock_init(&cfs_b->lock);
3492 cfs_b->quota = RUNTIME_INF;
3493 cfs_b->period = ns_to_ktime(default_cfs_period());
3495 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3496 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3497 cfs_b->period_timer.function = sched_cfs_period_timer;
3498 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3499 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3502 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3504 cfs_rq->runtime_enabled = 0;
3505 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3508 /* requires cfs_b->lock, may release to reprogram timer */
3509 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3512 * The timer may be active because we're trying to set a new bandwidth
3513 * period or because we're racing with the tear-down path
3514 * (timer_active==0 becomes visible before the hrtimer call-back
3515 * terminates). In either case we ensure that it's re-programmed
3517 while (unlikely(hrtimer_active(&cfs_b->period_timer)) &&
3518 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) {
3519 /* bounce the lock to allow do_sched_cfs_period_timer to run */
3520 raw_spin_unlock(&cfs_b->lock);
3522 raw_spin_lock(&cfs_b->lock);
3523 /* if someone else restarted the timer then we're done */
3524 if (cfs_b->timer_active)
3528 cfs_b->timer_active = 1;
3529 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3532 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3534 hrtimer_cancel(&cfs_b->period_timer);
3535 hrtimer_cancel(&cfs_b->slack_timer);
3538 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3540 struct cfs_rq *cfs_rq;
3542 for_each_leaf_cfs_rq(rq, cfs_rq) {
3543 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3545 if (!cfs_rq->runtime_enabled)
3549 * clock_task is not advancing so we just need to make sure
3550 * there's some valid quota amount
3552 cfs_rq->runtime_remaining = cfs_b->quota;
3553 if (cfs_rq_throttled(cfs_rq))
3554 unthrottle_cfs_rq(cfs_rq);
3558 #else /* CONFIG_CFS_BANDWIDTH */
3559 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3561 return rq_clock_task(rq_of(cfs_rq));
3564 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3565 unsigned long delta_exec) {}
3566 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3567 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3568 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3570 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3575 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3580 static inline int throttled_lb_pair(struct task_group *tg,
3581 int src_cpu, int dest_cpu)
3586 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3588 #ifdef CONFIG_FAIR_GROUP_SCHED
3589 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3592 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3596 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3597 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3599 #endif /* CONFIG_CFS_BANDWIDTH */
3601 /**************************************************
3602 * CFS operations on tasks:
3605 #ifdef CONFIG_SCHED_HRTICK
3606 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3608 struct sched_entity *se = &p->se;
3609 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3611 WARN_ON(task_rq(p) != rq);
3613 if (cfs_rq->nr_running > 1) {
3614 u64 slice = sched_slice(cfs_rq, se);
3615 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3616 s64 delta = slice - ran;
3625 * Don't schedule slices shorter than 10000ns, that just
3626 * doesn't make sense. Rely on vruntime for fairness.
3629 delta = max_t(s64, 10000LL, delta);
3631 hrtick_start(rq, delta);
3636 * called from enqueue/dequeue and updates the hrtick when the
3637 * current task is from our class and nr_running is low enough
3640 static void hrtick_update(struct rq *rq)
3642 struct task_struct *curr = rq->curr;
3644 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3647 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3648 hrtick_start_fair(rq, curr);
3650 #else /* !CONFIG_SCHED_HRTICK */
3652 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3656 static inline void hrtick_update(struct rq *rq)
3662 * The enqueue_task method is called before nr_running is
3663 * increased. Here we update the fair scheduling stats and
3664 * then put the task into the rbtree:
3667 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3669 struct cfs_rq *cfs_rq;
3670 struct sched_entity *se = &p->se;
3672 for_each_sched_entity(se) {
3675 cfs_rq = cfs_rq_of(se);
3676 enqueue_entity(cfs_rq, se, flags);
3679 * end evaluation on encountering a throttled cfs_rq
3681 * note: in the case of encountering a throttled cfs_rq we will
3682 * post the final h_nr_running increment below.
3684 if (cfs_rq_throttled(cfs_rq))
3686 cfs_rq->h_nr_running++;
3688 flags = ENQUEUE_WAKEUP;
3691 for_each_sched_entity(se) {
3692 cfs_rq = cfs_rq_of(se);
3693 cfs_rq->h_nr_running++;
3695 if (cfs_rq_throttled(cfs_rq))
3698 update_cfs_shares(cfs_rq);
3699 update_entity_load_avg(se, 1);
3703 update_rq_runnable_avg(rq, rq->nr_running);
3709 static void set_next_buddy(struct sched_entity *se);
3712 * The dequeue_task method is called before nr_running is
3713 * decreased. We remove the task from the rbtree and
3714 * update the fair scheduling stats:
3716 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3718 struct cfs_rq *cfs_rq;
3719 struct sched_entity *se = &p->se;
3720 int task_sleep = flags & DEQUEUE_SLEEP;
3722 for_each_sched_entity(se) {
3723 cfs_rq = cfs_rq_of(se);
3724 dequeue_entity(cfs_rq, se, flags);
3727 * end evaluation on encountering a throttled cfs_rq
3729 * note: in the case of encountering a throttled cfs_rq we will
3730 * post the final h_nr_running decrement below.
3732 if (cfs_rq_throttled(cfs_rq))
3734 cfs_rq->h_nr_running--;
3736 /* Don't dequeue parent if it has other entities besides us */
3737 if (cfs_rq->load.weight) {
3739 * Bias pick_next to pick a task from this cfs_rq, as
3740 * p is sleeping when it is within its sched_slice.
3742 if (task_sleep && parent_entity(se))
3743 set_next_buddy(parent_entity(se));
3745 /* avoid re-evaluating load for this entity */
3746 se = parent_entity(se);
3749 flags |= DEQUEUE_SLEEP;
3752 for_each_sched_entity(se) {
3753 cfs_rq = cfs_rq_of(se);
3754 cfs_rq->h_nr_running--;
3756 if (cfs_rq_throttled(cfs_rq))
3759 update_cfs_shares(cfs_rq);
3760 update_entity_load_avg(se, 1);
3765 update_rq_runnable_avg(rq, 1);
3771 /* Used instead of source_load when we know the type == 0 */
3772 static unsigned long weighted_cpuload(const int cpu)
3774 return cpu_rq(cpu)->cfs.runnable_load_avg;
3778 * Return a low guess at the load of a migration-source cpu weighted
3779 * according to the scheduling class and "nice" value.
3781 * We want to under-estimate the load of migration sources, to
3782 * balance conservatively.
3784 static unsigned long source_load(int cpu, int type)
3786 struct rq *rq = cpu_rq(cpu);
3787 unsigned long total = weighted_cpuload(cpu);
3789 if (type == 0 || !sched_feat(LB_BIAS))
3792 return min(rq->cpu_load[type-1], total);
3796 * Return a high guess at the load of a migration-target cpu weighted
3797 * according to the scheduling class and "nice" value.
3799 static unsigned long target_load(int cpu, int type)
3801 struct rq *rq = cpu_rq(cpu);
3802 unsigned long total = weighted_cpuload(cpu);
3804 if (type == 0 || !sched_feat(LB_BIAS))
3807 return max(rq->cpu_load[type-1], total);
3810 static unsigned long power_of(int cpu)
3812 return cpu_rq(cpu)->cpu_power;
3815 static unsigned long cpu_avg_load_per_task(int cpu)
3817 struct rq *rq = cpu_rq(cpu);
3818 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3819 unsigned long load_avg = rq->cfs.runnable_load_avg;
3822 return load_avg / nr_running;
3827 static void record_wakee(struct task_struct *p)
3830 * Rough decay (wiping) for cost saving, don't worry
3831 * about the boundary, really active task won't care
3834 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3835 current->wakee_flips = 0;
3836 current->wakee_flip_decay_ts = jiffies;
3839 if (current->last_wakee != p) {
3840 current->last_wakee = p;
3841 current->wakee_flips++;
3845 static void task_waking_fair(struct task_struct *p)
3847 struct sched_entity *se = &p->se;
3848 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3851 #ifndef CONFIG_64BIT
3852 u64 min_vruntime_copy;
3855 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3857 min_vruntime = cfs_rq->min_vruntime;
3858 } while (min_vruntime != min_vruntime_copy);
3860 min_vruntime = cfs_rq->min_vruntime;
3863 se->vruntime -= min_vruntime;
3867 #ifdef CONFIG_FAIR_GROUP_SCHED
3869 * effective_load() calculates the load change as seen from the root_task_group
3871 * Adding load to a group doesn't make a group heavier, but can cause movement
3872 * of group shares between cpus. Assuming the shares were perfectly aligned one
3873 * can calculate the shift in shares.
3875 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3876 * on this @cpu and results in a total addition (subtraction) of @wg to the
3877 * total group weight.
3879 * Given a runqueue weight distribution (rw_i) we can compute a shares
3880 * distribution (s_i) using:
3882 * s_i = rw_i / \Sum rw_j (1)
3884 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3885 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3886 * shares distribution (s_i):
3888 * rw_i = { 2, 4, 1, 0 }
3889 * s_i = { 2/7, 4/7, 1/7, 0 }
3891 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3892 * task used to run on and the CPU the waker is running on), we need to
3893 * compute the effect of waking a task on either CPU and, in case of a sync
3894 * wakeup, compute the effect of the current task going to sleep.
3896 * So for a change of @wl to the local @cpu with an overall group weight change
3897 * of @wl we can compute the new shares distribution (s'_i) using:
3899 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3901 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3902 * differences in waking a task to CPU 0. The additional task changes the
3903 * weight and shares distributions like:
3905 * rw'_i = { 3, 4, 1, 0 }
3906 * s'_i = { 3/8, 4/8, 1/8, 0 }
3908 * We can then compute the difference in effective weight by using:
3910 * dw_i = S * (s'_i - s_i) (3)
3912 * Where 'S' is the group weight as seen by its parent.
3914 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3915 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3916 * 4/7) times the weight of the group.
3918 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3920 struct sched_entity *se = tg->se[cpu];
3922 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3925 for_each_sched_entity(se) {
3931 * W = @wg + \Sum rw_j
3933 W = wg + calc_tg_weight(tg, se->my_q);
3938 w = se->my_q->load.weight + wl;
3941 * wl = S * s'_i; see (2)
3944 wl = (w * tg->shares) / W;
3949 * Per the above, wl is the new se->load.weight value; since
3950 * those are clipped to [MIN_SHARES, ...) do so now. See
3951 * calc_cfs_shares().
3953 if (wl < MIN_SHARES)
3957 * wl = dw_i = S * (s'_i - s_i); see (3)
3959 wl -= se->load.weight;
3962 * Recursively apply this logic to all parent groups to compute
3963 * the final effective load change on the root group. Since
3964 * only the @tg group gets extra weight, all parent groups can
3965 * only redistribute existing shares. @wl is the shift in shares
3966 * resulting from this level per the above.
3975 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3982 static int wake_wide(struct task_struct *p)
3984 int factor = this_cpu_read(sd_llc_size);
3987 * Yeah, it's the switching-frequency, could means many wakee or
3988 * rapidly switch, use factor here will just help to automatically
3989 * adjust the loose-degree, so bigger node will lead to more pull.
3991 if (p->wakee_flips > factor) {
3993 * wakee is somewhat hot, it needs certain amount of cpu
3994 * resource, so if waker is far more hot, prefer to leave
3997 if (current->wakee_flips > (factor * p->wakee_flips))
4004 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
4006 s64 this_load, load;
4007 int idx, this_cpu, prev_cpu;
4008 unsigned long tl_per_task;
4009 struct task_group *tg;
4010 unsigned long weight;
4014 * If we wake multiple tasks be careful to not bounce
4015 * ourselves around too much.
4021 this_cpu = smp_processor_id();
4022 prev_cpu = task_cpu(p);
4023 load = source_load(prev_cpu, idx);
4024 this_load = target_load(this_cpu, idx);
4027 * If sync wakeup then subtract the (maximum possible)
4028 * effect of the currently running task from the load
4029 * of the current CPU:
4032 tg = task_group(current);
4033 weight = current->se.load.weight;
4035 this_load += effective_load(tg, this_cpu, -weight, -weight);
4036 load += effective_load(tg, prev_cpu, 0, -weight);
4040 weight = p->se.load.weight;
4043 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4044 * due to the sync cause above having dropped this_load to 0, we'll
4045 * always have an imbalance, but there's really nothing you can do
4046 * about that, so that's good too.
4048 * Otherwise check if either cpus are near enough in load to allow this
4049 * task to be woken on this_cpu.
4051 if (this_load > 0) {
4052 s64 this_eff_load, prev_eff_load;
4054 this_eff_load = 100;
4055 this_eff_load *= power_of(prev_cpu);
4056 this_eff_load *= this_load +
4057 effective_load(tg, this_cpu, weight, weight);
4059 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4060 prev_eff_load *= power_of(this_cpu);
4061 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4063 balanced = this_eff_load <= prev_eff_load;
4068 * If the currently running task will sleep within
4069 * a reasonable amount of time then attract this newly
4072 if (sync && balanced)
4075 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4076 tl_per_task = cpu_avg_load_per_task(this_cpu);
4079 (this_load <= load &&
4080 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4082 * This domain has SD_WAKE_AFFINE and
4083 * p is cache cold in this domain, and
4084 * there is no bad imbalance.
4086 schedstat_inc(sd, ttwu_move_affine);
4087 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4095 * find_idlest_group finds and returns the least busy CPU group within the
4098 static struct sched_group *
4099 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4100 int this_cpu, int load_idx)
4102 struct sched_group *idlest = NULL, *group = sd->groups;
4103 unsigned long min_load = ULONG_MAX, this_load = 0;
4104 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4107 unsigned long load, avg_load;
4111 /* Skip over this group if it has no CPUs allowed */
4112 if (!cpumask_intersects(sched_group_cpus(group),
4113 tsk_cpus_allowed(p)))
4116 local_group = cpumask_test_cpu(this_cpu,
4117 sched_group_cpus(group));
4119 /* Tally up the load of all CPUs in the group */
4122 for_each_cpu(i, sched_group_cpus(group)) {
4123 /* Bias balancing toward cpus of our domain */
4125 load = source_load(i, load_idx);
4127 load = target_load(i, load_idx);
4132 /* Adjust by relative CPU power of the group */
4133 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4136 this_load = avg_load;
4137 } else if (avg_load < min_load) {
4138 min_load = avg_load;
4141 } while (group = group->next, group != sd->groups);
4143 if (!idlest || 100*this_load < imbalance*min_load)
4149 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4152 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4154 unsigned long load, min_load = ULONG_MAX;
4158 /* Traverse only the allowed CPUs */
4159 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4160 load = weighted_cpuload(i);
4162 if (load < min_load || (load == min_load && i == this_cpu)) {
4172 * Try and locate an idle CPU in the sched_domain.
4174 static int select_idle_sibling(struct task_struct *p, int target)
4176 struct sched_domain *sd;
4177 struct sched_group *sg;
4178 int i = task_cpu(p);
4180 if (idle_cpu(target))
4184 * If the prevous cpu is cache affine and idle, don't be stupid.
4186 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4190 * Otherwise, iterate the domains and find an elegible idle cpu.
4192 sd = rcu_dereference(per_cpu(sd_llc, target));
4193 for_each_lower_domain(sd) {
4196 if (!cpumask_intersects(sched_group_cpus(sg),
4197 tsk_cpus_allowed(p)))
4200 for_each_cpu(i, sched_group_cpus(sg)) {
4201 if (i == target || !idle_cpu(i))
4205 target = cpumask_first_and(sched_group_cpus(sg),
4206 tsk_cpus_allowed(p));
4210 } while (sg != sd->groups);
4217 * sched_balance_self: balance the current task (running on cpu) in domains
4218 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4221 * Balance, ie. select the least loaded group.
4223 * Returns the target CPU number, or the same CPU if no balancing is needed.
4225 * preempt must be disabled.
4228 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4230 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4231 int cpu = smp_processor_id();
4233 int want_affine = 0;
4234 int sync = wake_flags & WF_SYNC;
4236 if (p->nr_cpus_allowed == 1)
4239 if (sd_flag & SD_BALANCE_WAKE) {
4240 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4246 for_each_domain(cpu, tmp) {
4247 if (!(tmp->flags & SD_LOAD_BALANCE))
4251 * If both cpu and prev_cpu are part of this domain,
4252 * cpu is a valid SD_WAKE_AFFINE target.
4254 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4255 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4260 if (tmp->flags & sd_flag)
4265 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4268 new_cpu = select_idle_sibling(p, prev_cpu);
4273 int load_idx = sd->forkexec_idx;
4274 struct sched_group *group;
4277 if (!(sd->flags & sd_flag)) {
4282 if (sd_flag & SD_BALANCE_WAKE)
4283 load_idx = sd->wake_idx;
4285 group = find_idlest_group(sd, p, cpu, load_idx);
4291 new_cpu = find_idlest_cpu(group, p, cpu);
4292 if (new_cpu == -1 || new_cpu == cpu) {
4293 /* Now try balancing at a lower domain level of cpu */
4298 /* Now try balancing at a lower domain level of new_cpu */
4300 weight = sd->span_weight;
4302 for_each_domain(cpu, tmp) {
4303 if (weight <= tmp->span_weight)
4305 if (tmp->flags & sd_flag)
4308 /* while loop will break here if sd == NULL */
4317 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4318 * cfs_rq_of(p) references at time of call are still valid and identify the
4319 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4320 * other assumptions, including the state of rq->lock, should be made.
4323 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4325 struct sched_entity *se = &p->se;
4326 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4329 * Load tracking: accumulate removed load so that it can be processed
4330 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4331 * to blocked load iff they have a positive decay-count. It can never
4332 * be negative here since on-rq tasks have decay-count == 0.
4334 if (se->avg.decay_count) {
4335 se->avg.decay_count = -__synchronize_entity_decay(se);
4336 atomic_long_add(se->avg.load_avg_contrib,
4337 &cfs_rq->removed_load);
4340 #endif /* CONFIG_SMP */
4342 static unsigned long
4343 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4345 unsigned long gran = sysctl_sched_wakeup_granularity;
4348 * Since its curr running now, convert the gran from real-time
4349 * to virtual-time in his units.
4351 * By using 'se' instead of 'curr' we penalize light tasks, so
4352 * they get preempted easier. That is, if 'se' < 'curr' then
4353 * the resulting gran will be larger, therefore penalizing the
4354 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4355 * be smaller, again penalizing the lighter task.
4357 * This is especially important for buddies when the leftmost
4358 * task is higher priority than the buddy.
4360 return calc_delta_fair(gran, se);
4364 * Should 'se' preempt 'curr'.
4378 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4380 s64 gran, vdiff = curr->vruntime - se->vruntime;
4385 gran = wakeup_gran(curr, se);
4392 static void set_last_buddy(struct sched_entity *se)
4394 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4397 for_each_sched_entity(se)
4398 cfs_rq_of(se)->last = se;
4401 static void set_next_buddy(struct sched_entity *se)
4403 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4406 for_each_sched_entity(se)
4407 cfs_rq_of(se)->next = se;
4410 static void set_skip_buddy(struct sched_entity *se)
4412 for_each_sched_entity(se)
4413 cfs_rq_of(se)->skip = se;
4417 * Preempt the current task with a newly woken task if needed:
4419 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4421 struct task_struct *curr = rq->curr;
4422 struct sched_entity *se = &curr->se, *pse = &p->se;
4423 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4424 int scale = cfs_rq->nr_running >= sched_nr_latency;
4425 int next_buddy_marked = 0;
4427 if (unlikely(se == pse))
4431 * This is possible from callers such as move_task(), in which we
4432 * unconditionally check_prempt_curr() after an enqueue (which may have
4433 * lead to a throttle). This both saves work and prevents false
4434 * next-buddy nomination below.
4436 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4439 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4440 set_next_buddy(pse);
4441 next_buddy_marked = 1;
4445 * We can come here with TIF_NEED_RESCHED already set from new task
4448 * Note: this also catches the edge-case of curr being in a throttled
4449 * group (e.g. via set_curr_task), since update_curr() (in the
4450 * enqueue of curr) will have resulted in resched being set. This
4451 * prevents us from potentially nominating it as a false LAST_BUDDY
4454 if (test_tsk_need_resched(curr))
4457 /* Idle tasks are by definition preempted by non-idle tasks. */
4458 if (unlikely(curr->policy == SCHED_IDLE) &&
4459 likely(p->policy != SCHED_IDLE))
4463 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4464 * is driven by the tick):
4466 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4469 find_matching_se(&se, &pse);
4470 update_curr(cfs_rq_of(se));
4472 if (wakeup_preempt_entity(se, pse) == 1) {
4474 * Bias pick_next to pick the sched entity that is
4475 * triggering this preemption.
4477 if (!next_buddy_marked)
4478 set_next_buddy(pse);
4487 * Only set the backward buddy when the current task is still
4488 * on the rq. This can happen when a wakeup gets interleaved
4489 * with schedule on the ->pre_schedule() or idle_balance()
4490 * point, either of which can * drop the rq lock.
4492 * Also, during early boot the idle thread is in the fair class,
4493 * for obvious reasons its a bad idea to schedule back to it.
4495 if (unlikely(!se->on_rq || curr == rq->idle))
4498 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4502 static struct task_struct *pick_next_task_fair(struct rq *rq)
4504 struct task_struct *p;
4505 struct cfs_rq *cfs_rq = &rq->cfs;
4506 struct sched_entity *se;
4508 if (!cfs_rq->nr_running)
4512 se = pick_next_entity(cfs_rq);
4513 set_next_entity(cfs_rq, se);
4514 cfs_rq = group_cfs_rq(se);
4518 if (hrtick_enabled(rq))
4519 hrtick_start_fair(rq, p);
4525 * Account for a descheduled task:
4527 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4529 struct sched_entity *se = &prev->se;
4530 struct cfs_rq *cfs_rq;
4532 for_each_sched_entity(se) {
4533 cfs_rq = cfs_rq_of(se);
4534 put_prev_entity(cfs_rq, se);
4539 * sched_yield() is very simple
4541 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4543 static void yield_task_fair(struct rq *rq)
4545 struct task_struct *curr = rq->curr;
4546 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4547 struct sched_entity *se = &curr->se;
4550 * Are we the only task in the tree?
4552 if (unlikely(rq->nr_running == 1))
4555 clear_buddies(cfs_rq, se);
4557 if (curr->policy != SCHED_BATCH) {
4558 update_rq_clock(rq);
4560 * Update run-time statistics of the 'current'.
4562 update_curr(cfs_rq);
4564 * Tell update_rq_clock() that we've just updated,
4565 * so we don't do microscopic update in schedule()
4566 * and double the fastpath cost.
4568 rq->skip_clock_update = 1;
4574 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4576 struct sched_entity *se = &p->se;
4578 /* throttled hierarchies are not runnable */
4579 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4582 /* Tell the scheduler that we'd really like pse to run next. */
4585 yield_task_fair(rq);
4591 /**************************************************
4592 * Fair scheduling class load-balancing methods.
4596 * The purpose of load-balancing is to achieve the same basic fairness the
4597 * per-cpu scheduler provides, namely provide a proportional amount of compute
4598 * time to each task. This is expressed in the following equation:
4600 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4602 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4603 * W_i,0 is defined as:
4605 * W_i,0 = \Sum_j w_i,j (2)
4607 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4608 * is derived from the nice value as per prio_to_weight[].
4610 * The weight average is an exponential decay average of the instantaneous
4613 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4615 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4616 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4617 * can also include other factors [XXX].
4619 * To achieve this balance we define a measure of imbalance which follows
4620 * directly from (1):
4622 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4624 * We them move tasks around to minimize the imbalance. In the continuous
4625 * function space it is obvious this converges, in the discrete case we get
4626 * a few fun cases generally called infeasible weight scenarios.
4629 * - infeasible weights;
4630 * - local vs global optima in the discrete case. ]
4635 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4636 * for all i,j solution, we create a tree of cpus that follows the hardware
4637 * topology where each level pairs two lower groups (or better). This results
4638 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4639 * tree to only the first of the previous level and we decrease the frequency
4640 * of load-balance at each level inv. proportional to the number of cpus in
4646 * \Sum { --- * --- * 2^i } = O(n) (5)
4648 * `- size of each group
4649 * | | `- number of cpus doing load-balance
4651 * `- sum over all levels
4653 * Coupled with a limit on how many tasks we can migrate every balance pass,
4654 * this makes (5) the runtime complexity of the balancer.
4656 * An important property here is that each CPU is still (indirectly) connected
4657 * to every other cpu in at most O(log n) steps:
4659 * The adjacency matrix of the resulting graph is given by:
4662 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4665 * And you'll find that:
4667 * A^(log_2 n)_i,j != 0 for all i,j (7)
4669 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4670 * The task movement gives a factor of O(m), giving a convergence complexity
4673 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4678 * In order to avoid CPUs going idle while there's still work to do, new idle
4679 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4680 * tree itself instead of relying on other CPUs to bring it work.
4682 * This adds some complexity to both (5) and (8) but it reduces the total idle
4690 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4693 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4698 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4700 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4702 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4705 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4706 * rewrite all of this once again.]
4709 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4711 enum fbq_type { regular, remote, all };
4713 #define LBF_ALL_PINNED 0x01
4714 #define LBF_NEED_BREAK 0x02
4715 #define LBF_DST_PINNED 0x04
4716 #define LBF_SOME_PINNED 0x08
4719 struct sched_domain *sd;
4727 struct cpumask *dst_grpmask;
4729 enum cpu_idle_type idle;
4731 /* The set of CPUs under consideration for load-balancing */
4732 struct cpumask *cpus;
4737 unsigned int loop_break;
4738 unsigned int loop_max;
4740 enum fbq_type fbq_type;
4744 * move_task - move a task from one runqueue to another runqueue.
4745 * Both runqueues must be locked.
4747 static void move_task(struct task_struct *p, struct lb_env *env)
4749 deactivate_task(env->src_rq, p, 0);
4750 set_task_cpu(p, env->dst_cpu);
4751 activate_task(env->dst_rq, p, 0);
4752 check_preempt_curr(env->dst_rq, p, 0);
4756 * Is this task likely cache-hot:
4759 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4763 if (p->sched_class != &fair_sched_class)
4766 if (unlikely(p->policy == SCHED_IDLE))
4770 * Buddy candidates are cache hot:
4772 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4773 (&p->se == cfs_rq_of(&p->se)->next ||
4774 &p->se == cfs_rq_of(&p->se)->last))
4777 if (sysctl_sched_migration_cost == -1)
4779 if (sysctl_sched_migration_cost == 0)
4782 delta = now - p->se.exec_start;
4784 return delta < (s64)sysctl_sched_migration_cost;
4787 #ifdef CONFIG_NUMA_BALANCING
4788 /* Returns true if the destination node has incurred more faults */
4789 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4791 int src_nid, dst_nid;
4793 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4794 !(env->sd->flags & SD_NUMA)) {
4798 src_nid = cpu_to_node(env->src_cpu);
4799 dst_nid = cpu_to_node(env->dst_cpu);
4801 if (src_nid == dst_nid)
4804 /* Always encourage migration to the preferred node. */
4805 if (dst_nid == p->numa_preferred_nid)
4808 /* If both task and group weight improve, this move is a winner. */
4809 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4810 group_weight(p, dst_nid) > group_weight(p, src_nid))
4817 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4819 int src_nid, dst_nid;
4821 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4824 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4827 src_nid = cpu_to_node(env->src_cpu);
4828 dst_nid = cpu_to_node(env->dst_cpu);
4830 if (src_nid == dst_nid)
4833 /* Migrating away from the preferred node is always bad. */
4834 if (src_nid == p->numa_preferred_nid)
4837 /* If either task or group weight get worse, don't do it. */
4838 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4839 group_weight(p, dst_nid) < group_weight(p, src_nid))
4846 static inline bool migrate_improves_locality(struct task_struct *p,
4852 static inline bool migrate_degrades_locality(struct task_struct *p,
4860 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4863 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4865 int tsk_cache_hot = 0;
4867 * We do not migrate tasks that are:
4868 * 1) throttled_lb_pair, or
4869 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4870 * 3) running (obviously), or
4871 * 4) are cache-hot on their current CPU.
4873 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4876 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4879 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4881 env->flags |= LBF_SOME_PINNED;
4884 * Remember if this task can be migrated to any other cpu in
4885 * our sched_group. We may want to revisit it if we couldn't
4886 * meet load balance goals by pulling other tasks on src_cpu.
4888 * Also avoid computing new_dst_cpu if we have already computed
4889 * one in current iteration.
4891 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4894 /* Prevent to re-select dst_cpu via env's cpus */
4895 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4896 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4897 env->flags |= LBF_DST_PINNED;
4898 env->new_dst_cpu = cpu;
4906 /* Record that we found atleast one task that could run on dst_cpu */
4907 env->flags &= ~LBF_ALL_PINNED;
4909 if (task_running(env->src_rq, p)) {
4910 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4915 * Aggressive migration if:
4916 * 1) destination numa is preferred
4917 * 2) task is cache cold, or
4918 * 3) too many balance attempts have failed.
4920 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4922 tsk_cache_hot = migrate_degrades_locality(p, env);
4924 if (migrate_improves_locality(p, env)) {
4925 #ifdef CONFIG_SCHEDSTATS
4926 if (tsk_cache_hot) {
4927 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4928 schedstat_inc(p, se.statistics.nr_forced_migrations);
4934 if (!tsk_cache_hot ||
4935 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4937 if (tsk_cache_hot) {
4938 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4939 schedstat_inc(p, se.statistics.nr_forced_migrations);
4945 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4950 * move_one_task tries to move exactly one task from busiest to this_rq, as
4951 * part of active balancing operations within "domain".
4952 * Returns 1 if successful and 0 otherwise.
4954 * Called with both runqueues locked.
4956 static int move_one_task(struct lb_env *env)
4958 struct task_struct *p, *n;
4960 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4961 if (!can_migrate_task(p, env))
4966 * Right now, this is only the second place move_task()
4967 * is called, so we can safely collect move_task()
4968 * stats here rather than inside move_task().
4970 schedstat_inc(env->sd, lb_gained[env->idle]);
4976 static const unsigned int sched_nr_migrate_break = 32;
4979 * move_tasks tries to move up to imbalance weighted load from busiest to
4980 * this_rq, as part of a balancing operation within domain "sd".
4981 * Returns 1 if successful and 0 otherwise.
4983 * Called with both runqueues locked.
4985 static int move_tasks(struct lb_env *env)
4987 struct list_head *tasks = &env->src_rq->cfs_tasks;
4988 struct task_struct *p;
4992 if (env->imbalance <= 0)
4995 while (!list_empty(tasks)) {
4996 p = list_first_entry(tasks, struct task_struct, se.group_node);
4999 /* We've more or less seen every task there is, call it quits */
5000 if (env->loop > env->loop_max)
5003 /* take a breather every nr_migrate tasks */
5004 if (env->loop > env->loop_break) {
5005 env->loop_break += sched_nr_migrate_break;
5006 env->flags |= LBF_NEED_BREAK;
5010 if (!can_migrate_task(p, env))
5013 load = task_h_load(p);
5015 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
5018 if ((load / 2) > env->imbalance)
5023 env->imbalance -= load;
5025 #ifdef CONFIG_PREEMPT
5027 * NEWIDLE balancing is a source of latency, so preemptible
5028 * kernels will stop after the first task is pulled to minimize
5029 * the critical section.
5031 if (env->idle == CPU_NEWLY_IDLE)
5036 * We only want to steal up to the prescribed amount of
5039 if (env->imbalance <= 0)
5044 list_move_tail(&p->se.group_node, tasks);
5048 * Right now, this is one of only two places move_task() is called,
5049 * so we can safely collect move_task() stats here rather than
5050 * inside move_task().
5052 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5057 #ifdef CONFIG_FAIR_GROUP_SCHED
5059 * update tg->load_weight by folding this cpu's load_avg
5061 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5063 struct sched_entity *se = tg->se[cpu];
5064 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5066 /* throttled entities do not contribute to load */
5067 if (throttled_hierarchy(cfs_rq))
5070 update_cfs_rq_blocked_load(cfs_rq, 1);
5073 update_entity_load_avg(se, 1);
5075 * We pivot on our runnable average having decayed to zero for
5076 * list removal. This generally implies that all our children
5077 * have also been removed (modulo rounding error or bandwidth
5078 * control); however, such cases are rare and we can fix these
5081 * TODO: fix up out-of-order children on enqueue.
5083 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5084 list_del_leaf_cfs_rq(cfs_rq);
5086 struct rq *rq = rq_of(cfs_rq);
5087 update_rq_runnable_avg(rq, rq->nr_running);
5091 static void update_blocked_averages(int cpu)
5093 struct rq *rq = cpu_rq(cpu);
5094 struct cfs_rq *cfs_rq;
5095 unsigned long flags;
5097 raw_spin_lock_irqsave(&rq->lock, flags);
5098 update_rq_clock(rq);
5100 * Iterates the task_group tree in a bottom up fashion, see
5101 * list_add_leaf_cfs_rq() for details.
5103 for_each_leaf_cfs_rq(rq, cfs_rq) {
5105 * Note: We may want to consider periodically releasing
5106 * rq->lock about these updates so that creating many task
5107 * groups does not result in continually extending hold time.
5109 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5112 raw_spin_unlock_irqrestore(&rq->lock, flags);
5116 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5117 * This needs to be done in a top-down fashion because the load of a child
5118 * group is a fraction of its parents load.
5120 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5122 struct rq *rq = rq_of(cfs_rq);
5123 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5124 unsigned long now = jiffies;
5127 if (cfs_rq->last_h_load_update == now)
5130 cfs_rq->h_load_next = NULL;
5131 for_each_sched_entity(se) {
5132 cfs_rq = cfs_rq_of(se);
5133 cfs_rq->h_load_next = se;
5134 if (cfs_rq->last_h_load_update == now)
5139 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5140 cfs_rq->last_h_load_update = now;
5143 while ((se = cfs_rq->h_load_next) != NULL) {
5144 load = cfs_rq->h_load;
5145 load = div64_ul(load * se->avg.load_avg_contrib,
5146 cfs_rq->runnable_load_avg + 1);
5147 cfs_rq = group_cfs_rq(se);
5148 cfs_rq->h_load = load;
5149 cfs_rq->last_h_load_update = now;
5153 static unsigned long task_h_load(struct task_struct *p)
5155 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5157 update_cfs_rq_h_load(cfs_rq);
5158 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5159 cfs_rq->runnable_load_avg + 1);
5162 static inline void update_blocked_averages(int cpu)
5166 static unsigned long task_h_load(struct task_struct *p)
5168 return p->se.avg.load_avg_contrib;
5172 /********** Helpers for find_busiest_group ************************/
5174 * sg_lb_stats - stats of a sched_group required for load_balancing
5176 struct sg_lb_stats {
5177 unsigned long avg_load; /*Avg load across the CPUs of the group */
5178 unsigned long group_load; /* Total load over the CPUs of the group */
5179 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5180 unsigned long load_per_task;
5181 unsigned long group_power;
5182 unsigned int sum_nr_running; /* Nr tasks running in the group */
5183 unsigned int group_capacity;
5184 unsigned int idle_cpus;
5185 unsigned int group_weight;
5186 int group_imb; /* Is there an imbalance in the group ? */
5187 int group_has_capacity; /* Is there extra capacity in the group? */
5188 #ifdef CONFIG_NUMA_BALANCING
5189 unsigned int nr_numa_running;
5190 unsigned int nr_preferred_running;
5195 * sd_lb_stats - Structure to store the statistics of a sched_domain
5196 * during load balancing.
5198 struct sd_lb_stats {
5199 struct sched_group *busiest; /* Busiest group in this sd */
5200 struct sched_group *local; /* Local group in this sd */
5201 unsigned long total_load; /* Total load of all groups in sd */
5202 unsigned long total_pwr; /* Total power of all groups in sd */
5203 unsigned long avg_load; /* Average load across all groups in sd */
5205 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5206 struct sg_lb_stats local_stat; /* Statistics of the local group */
5209 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5212 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5213 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5214 * We must however clear busiest_stat::avg_load because
5215 * update_sd_pick_busiest() reads this before assignment.
5217 *sds = (struct sd_lb_stats){
5229 * get_sd_load_idx - Obtain the load index for a given sched domain.
5230 * @sd: The sched_domain whose load_idx is to be obtained.
5231 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5233 * Return: The load index.
5235 static inline int get_sd_load_idx(struct sched_domain *sd,
5236 enum cpu_idle_type idle)
5242 load_idx = sd->busy_idx;
5245 case CPU_NEWLY_IDLE:
5246 load_idx = sd->newidle_idx;
5249 load_idx = sd->idle_idx;
5256 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5258 return SCHED_POWER_SCALE;
5261 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5263 return default_scale_freq_power(sd, cpu);
5266 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5268 unsigned long weight = sd->span_weight;
5269 unsigned long smt_gain = sd->smt_gain;
5276 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5278 return default_scale_smt_power(sd, cpu);
5281 static unsigned long scale_rt_power(int cpu)
5283 struct rq *rq = cpu_rq(cpu);
5284 u64 total, available, age_stamp, avg;
5287 * Since we're reading these variables without serialization make sure
5288 * we read them once before doing sanity checks on them.
5290 age_stamp = ACCESS_ONCE(rq->age_stamp);
5291 avg = ACCESS_ONCE(rq->rt_avg);
5293 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5295 if (unlikely(total < avg)) {
5296 /* Ensures that power won't end up being negative */
5299 available = total - avg;
5302 if (unlikely((s64)total < SCHED_POWER_SCALE))
5303 total = SCHED_POWER_SCALE;
5305 total >>= SCHED_POWER_SHIFT;
5307 return div_u64(available, total);
5310 static void update_cpu_power(struct sched_domain *sd, int cpu)
5312 unsigned long weight = sd->span_weight;
5313 unsigned long power = SCHED_POWER_SCALE;
5314 struct sched_group *sdg = sd->groups;
5316 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5317 if (sched_feat(ARCH_POWER))
5318 power *= arch_scale_smt_power(sd, cpu);
5320 power *= default_scale_smt_power(sd, cpu);
5322 power >>= SCHED_POWER_SHIFT;
5325 sdg->sgp->power_orig = power;
5327 if (sched_feat(ARCH_POWER))
5328 power *= arch_scale_freq_power(sd, cpu);
5330 power *= default_scale_freq_power(sd, cpu);
5332 power >>= SCHED_POWER_SHIFT;
5334 power *= scale_rt_power(cpu);
5335 power >>= SCHED_POWER_SHIFT;
5340 cpu_rq(cpu)->cpu_power = power;
5341 sdg->sgp->power = power;
5344 void update_group_power(struct sched_domain *sd, int cpu)
5346 struct sched_domain *child = sd->child;
5347 struct sched_group *group, *sdg = sd->groups;
5348 unsigned long power, power_orig;
5349 unsigned long interval;
5351 interval = msecs_to_jiffies(sd->balance_interval);
5352 interval = clamp(interval, 1UL, max_load_balance_interval);
5353 sdg->sgp->next_update = jiffies + interval;
5356 update_cpu_power(sd, cpu);
5360 power_orig = power = 0;
5362 if (child->flags & SD_OVERLAP) {
5364 * SD_OVERLAP domains cannot assume that child groups
5365 * span the current group.
5368 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5369 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5371 power_orig += sg->sgp->power_orig;
5372 power += sg->sgp->power;
5376 * !SD_OVERLAP domains can assume that child groups
5377 * span the current group.
5380 group = child->groups;
5382 power_orig += group->sgp->power_orig;
5383 power += group->sgp->power;
5384 group = group->next;
5385 } while (group != child->groups);
5388 sdg->sgp->power_orig = power_orig;
5389 sdg->sgp->power = power;
5393 * Try and fix up capacity for tiny siblings, this is needed when
5394 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5395 * which on its own isn't powerful enough.
5397 * See update_sd_pick_busiest() and check_asym_packing().
5400 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5403 * Only siblings can have significantly less than SCHED_POWER_SCALE
5405 if (!(sd->flags & SD_SHARE_CPUPOWER))
5409 * If ~90% of the cpu_power is still there, we're good.
5411 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5418 * Group imbalance indicates (and tries to solve) the problem where balancing
5419 * groups is inadequate due to tsk_cpus_allowed() constraints.
5421 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5422 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5425 * { 0 1 2 3 } { 4 5 6 7 }
5428 * If we were to balance group-wise we'd place two tasks in the first group and
5429 * two tasks in the second group. Clearly this is undesired as it will overload
5430 * cpu 3 and leave one of the cpus in the second group unused.
5432 * The current solution to this issue is detecting the skew in the first group
5433 * by noticing the lower domain failed to reach balance and had difficulty
5434 * moving tasks due to affinity constraints.
5436 * When this is so detected; this group becomes a candidate for busiest; see
5437 * update_sd_pick_busiest(). And calculate_imbalance() and
5438 * find_busiest_group() avoid some of the usual balance conditions to allow it
5439 * to create an effective group imbalance.
5441 * This is a somewhat tricky proposition since the next run might not find the
5442 * group imbalance and decide the groups need to be balanced again. A most
5443 * subtle and fragile situation.
5446 static inline int sg_imbalanced(struct sched_group *group)
5448 return group->sgp->imbalance;
5452 * Compute the group capacity.
5454 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5455 * first dividing out the smt factor and computing the actual number of cores
5456 * and limit power unit capacity with that.
5458 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5460 unsigned int capacity, smt, cpus;
5461 unsigned int power, power_orig;
5463 power = group->sgp->power;
5464 power_orig = group->sgp->power_orig;
5465 cpus = group->group_weight;
5467 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5468 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5469 capacity = cpus / smt; /* cores */
5471 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5473 capacity = fix_small_capacity(env->sd, group);
5479 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5480 * @env: The load balancing environment.
5481 * @group: sched_group whose statistics are to be updated.
5482 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5483 * @local_group: Does group contain this_cpu.
5484 * @sgs: variable to hold the statistics for this group.
5486 static inline void update_sg_lb_stats(struct lb_env *env,
5487 struct sched_group *group, int load_idx,
5488 int local_group, struct sg_lb_stats *sgs)
5490 unsigned long nr_running;
5494 memset(sgs, 0, sizeof(*sgs));
5496 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5497 struct rq *rq = cpu_rq(i);
5499 nr_running = rq->nr_running;
5501 /* Bias balancing toward cpus of our domain */
5503 load = target_load(i, load_idx);
5505 load = source_load(i, load_idx);
5507 sgs->group_load += load;
5508 sgs->sum_nr_running += nr_running;
5509 #ifdef CONFIG_NUMA_BALANCING
5510 sgs->nr_numa_running += rq->nr_numa_running;
5511 sgs->nr_preferred_running += rq->nr_preferred_running;
5513 sgs->sum_weighted_load += weighted_cpuload(i);
5518 /* Adjust by relative CPU power of the group */
5519 sgs->group_power = group->sgp->power;
5520 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5522 if (sgs->sum_nr_running)
5523 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5525 sgs->group_weight = group->group_weight;
5527 sgs->group_imb = sg_imbalanced(group);
5528 sgs->group_capacity = sg_capacity(env, group);
5530 if (sgs->group_capacity > sgs->sum_nr_running)
5531 sgs->group_has_capacity = 1;
5535 * update_sd_pick_busiest - return 1 on busiest group
5536 * @env: The load balancing environment.
5537 * @sds: sched_domain statistics
5538 * @sg: sched_group candidate to be checked for being the busiest
5539 * @sgs: sched_group statistics
5541 * Determine if @sg is a busier group than the previously selected
5544 * Return: %true if @sg is a busier group than the previously selected
5545 * busiest group. %false otherwise.
5547 static bool update_sd_pick_busiest(struct lb_env *env,
5548 struct sd_lb_stats *sds,
5549 struct sched_group *sg,
5550 struct sg_lb_stats *sgs)
5552 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5555 if (sgs->sum_nr_running > sgs->group_capacity)
5562 * ASYM_PACKING needs to move all the work to the lowest
5563 * numbered CPUs in the group, therefore mark all groups
5564 * higher than ourself as busy.
5566 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5567 env->dst_cpu < group_first_cpu(sg)) {
5571 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5578 #ifdef CONFIG_NUMA_BALANCING
5579 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5581 if (sgs->sum_nr_running > sgs->nr_numa_running)
5583 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5588 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5590 if (rq->nr_running > rq->nr_numa_running)
5592 if (rq->nr_running > rq->nr_preferred_running)
5597 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5602 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5606 #endif /* CONFIG_NUMA_BALANCING */
5609 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5610 * @env: The load balancing environment.
5611 * @sds: variable to hold the statistics for this sched_domain.
5613 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5615 struct sched_domain *child = env->sd->child;
5616 struct sched_group *sg = env->sd->groups;
5617 struct sg_lb_stats tmp_sgs;
5618 int load_idx, prefer_sibling = 0;
5620 if (child && child->flags & SD_PREFER_SIBLING)
5623 load_idx = get_sd_load_idx(env->sd, env->idle);
5626 struct sg_lb_stats *sgs = &tmp_sgs;
5629 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5632 sgs = &sds->local_stat;
5634 if (env->idle != CPU_NEWLY_IDLE ||
5635 time_after_eq(jiffies, sg->sgp->next_update))
5636 update_group_power(env->sd, env->dst_cpu);
5639 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5645 * In case the child domain prefers tasks go to siblings
5646 * first, lower the sg capacity to one so that we'll try
5647 * and move all the excess tasks away. We lower the capacity
5648 * of a group only if the local group has the capacity to fit
5649 * these excess tasks, i.e. nr_running < group_capacity. The
5650 * extra check prevents the case where you always pull from the
5651 * heaviest group when it is already under-utilized (possible
5652 * with a large weight task outweighs the tasks on the system).
5654 if (prefer_sibling && sds->local &&
5655 sds->local_stat.group_has_capacity)
5656 sgs->group_capacity = min(sgs->group_capacity, 1U);
5658 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5660 sds->busiest_stat = *sgs;
5664 /* Now, start updating sd_lb_stats */
5665 sds->total_load += sgs->group_load;
5666 sds->total_pwr += sgs->group_power;
5669 } while (sg != env->sd->groups);
5671 if (env->sd->flags & SD_NUMA)
5672 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5676 * check_asym_packing - Check to see if the group is packed into the
5679 * This is primarily intended to used at the sibling level. Some
5680 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5681 * case of POWER7, it can move to lower SMT modes only when higher
5682 * threads are idle. When in lower SMT modes, the threads will
5683 * perform better since they share less core resources. Hence when we
5684 * have idle threads, we want them to be the higher ones.
5686 * This packing function is run on idle threads. It checks to see if
5687 * the busiest CPU in this domain (core in the P7 case) has a higher
5688 * CPU number than the packing function is being run on. Here we are
5689 * assuming lower CPU number will be equivalent to lower a SMT thread
5692 * Return: 1 when packing is required and a task should be moved to
5693 * this CPU. The amount of the imbalance is returned in *imbalance.
5695 * @env: The load balancing environment.
5696 * @sds: Statistics of the sched_domain which is to be packed
5698 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5702 if (!(env->sd->flags & SD_ASYM_PACKING))
5708 busiest_cpu = group_first_cpu(sds->busiest);
5709 if (env->dst_cpu > busiest_cpu)
5712 env->imbalance = DIV_ROUND_CLOSEST(
5713 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5720 * fix_small_imbalance - Calculate the minor imbalance that exists
5721 * amongst the groups of a sched_domain, during
5723 * @env: The load balancing environment.
5724 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5727 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5729 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5730 unsigned int imbn = 2;
5731 unsigned long scaled_busy_load_per_task;
5732 struct sg_lb_stats *local, *busiest;
5734 local = &sds->local_stat;
5735 busiest = &sds->busiest_stat;
5737 if (!local->sum_nr_running)
5738 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5739 else if (busiest->load_per_task > local->load_per_task)
5742 scaled_busy_load_per_task =
5743 (busiest->load_per_task * SCHED_POWER_SCALE) /
5744 busiest->group_power;
5746 if (busiest->avg_load + scaled_busy_load_per_task >=
5747 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5748 env->imbalance = busiest->load_per_task;
5753 * OK, we don't have enough imbalance to justify moving tasks,
5754 * however we may be able to increase total CPU power used by
5758 pwr_now += busiest->group_power *
5759 min(busiest->load_per_task, busiest->avg_load);
5760 pwr_now += local->group_power *
5761 min(local->load_per_task, local->avg_load);
5762 pwr_now /= SCHED_POWER_SCALE;
5764 /* Amount of load we'd subtract */
5765 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5766 busiest->group_power;
5767 if (busiest->avg_load > tmp) {
5768 pwr_move += busiest->group_power *
5769 min(busiest->load_per_task,
5770 busiest->avg_load - tmp);
5773 /* Amount of load we'd add */
5774 if (busiest->avg_load * busiest->group_power <
5775 busiest->load_per_task * SCHED_POWER_SCALE) {
5776 tmp = (busiest->avg_load * busiest->group_power) /
5779 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5782 pwr_move += local->group_power *
5783 min(local->load_per_task, local->avg_load + tmp);
5784 pwr_move /= SCHED_POWER_SCALE;
5786 /* Move if we gain throughput */
5787 if (pwr_move > pwr_now)
5788 env->imbalance = busiest->load_per_task;
5792 * calculate_imbalance - Calculate the amount of imbalance present within the
5793 * groups of a given sched_domain during load balance.
5794 * @env: load balance environment
5795 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5797 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5799 unsigned long max_pull, load_above_capacity = ~0UL;
5800 struct sg_lb_stats *local, *busiest;
5802 local = &sds->local_stat;
5803 busiest = &sds->busiest_stat;
5805 if (busiest->group_imb) {
5807 * In the group_imb case we cannot rely on group-wide averages
5808 * to ensure cpu-load equilibrium, look at wider averages. XXX
5810 busiest->load_per_task =
5811 min(busiest->load_per_task, sds->avg_load);
5815 * In the presence of smp nice balancing, certain scenarios can have
5816 * max load less than avg load(as we skip the groups at or below
5817 * its cpu_power, while calculating max_load..)
5819 if (busiest->avg_load <= sds->avg_load ||
5820 local->avg_load >= sds->avg_load) {
5822 return fix_small_imbalance(env, sds);
5825 if (!busiest->group_imb) {
5827 * Don't want to pull so many tasks that a group would go idle.
5828 * Except of course for the group_imb case, since then we might
5829 * have to drop below capacity to reach cpu-load equilibrium.
5831 load_above_capacity =
5832 (busiest->sum_nr_running - busiest->group_capacity);
5834 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5835 load_above_capacity /= busiest->group_power;
5839 * We're trying to get all the cpus to the average_load, so we don't
5840 * want to push ourselves above the average load, nor do we wish to
5841 * reduce the max loaded cpu below the average load. At the same time,
5842 * we also don't want to reduce the group load below the group capacity
5843 * (so that we can implement power-savings policies etc). Thus we look
5844 * for the minimum possible imbalance.
5846 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5848 /* How much load to actually move to equalise the imbalance */
5849 env->imbalance = min(
5850 max_pull * busiest->group_power,
5851 (sds->avg_load - local->avg_load) * local->group_power
5852 ) / SCHED_POWER_SCALE;
5855 * if *imbalance is less than the average load per runnable task
5856 * there is no guarantee that any tasks will be moved so we'll have
5857 * a think about bumping its value to force at least one task to be
5860 if (env->imbalance < busiest->load_per_task)
5861 return fix_small_imbalance(env, sds);
5864 /******* find_busiest_group() helpers end here *********************/
5867 * find_busiest_group - Returns the busiest group within the sched_domain
5868 * if there is an imbalance. If there isn't an imbalance, and
5869 * the user has opted for power-savings, it returns a group whose
5870 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5871 * such a group exists.
5873 * Also calculates the amount of weighted load which should be moved
5874 * to restore balance.
5876 * @env: The load balancing environment.
5878 * Return: - The busiest group if imbalance exists.
5879 * - If no imbalance and user has opted for power-savings balance,
5880 * return the least loaded group whose CPUs can be
5881 * put to idle by rebalancing its tasks onto our group.
5883 static struct sched_group *find_busiest_group(struct lb_env *env)
5885 struct sg_lb_stats *local, *busiest;
5886 struct sd_lb_stats sds;
5888 init_sd_lb_stats(&sds);
5891 * Compute the various statistics relavent for load balancing at
5894 update_sd_lb_stats(env, &sds);
5895 local = &sds.local_stat;
5896 busiest = &sds.busiest_stat;
5898 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5899 check_asym_packing(env, &sds))
5902 /* There is no busy sibling group to pull tasks from */
5903 if (!sds.busiest || busiest->sum_nr_running == 0)
5906 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5909 * If the busiest group is imbalanced the below checks don't
5910 * work because they assume all things are equal, which typically
5911 * isn't true due to cpus_allowed constraints and the like.
5913 if (busiest->group_imb)
5916 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5917 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5918 !busiest->group_has_capacity)
5922 * If the local group is more busy than the selected busiest group
5923 * don't try and pull any tasks.
5925 if (local->avg_load >= busiest->avg_load)
5929 * Don't pull any tasks if this group is already above the domain
5932 if (local->avg_load >= sds.avg_load)
5935 if (env->idle == CPU_IDLE) {
5937 * This cpu is idle. If the busiest group load doesn't
5938 * have more tasks than the number of available cpu's and
5939 * there is no imbalance between this and busiest group
5940 * wrt to idle cpu's, it is balanced.
5942 if ((local->idle_cpus < busiest->idle_cpus) &&
5943 busiest->sum_nr_running <= busiest->group_weight)
5947 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5948 * imbalance_pct to be conservative.
5950 if (100 * busiest->avg_load <=
5951 env->sd->imbalance_pct * local->avg_load)
5956 /* Looks like there is an imbalance. Compute it */
5957 calculate_imbalance(env, &sds);
5966 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5968 static struct rq *find_busiest_queue(struct lb_env *env,
5969 struct sched_group *group)
5971 struct rq *busiest = NULL, *rq;
5972 unsigned long busiest_load = 0, busiest_power = 1;
5975 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5976 unsigned long power, capacity, wl;
5980 rt = fbq_classify_rq(rq);
5983 * We classify groups/runqueues into three groups:
5984 * - regular: there are !numa tasks
5985 * - remote: there are numa tasks that run on the 'wrong' node
5986 * - all: there is no distinction
5988 * In order to avoid migrating ideally placed numa tasks,
5989 * ignore those when there's better options.
5991 * If we ignore the actual busiest queue to migrate another
5992 * task, the next balance pass can still reduce the busiest
5993 * queue by moving tasks around inside the node.
5995 * If we cannot move enough load due to this classification
5996 * the next pass will adjust the group classification and
5997 * allow migration of more tasks.
5999 * Both cases only affect the total convergence complexity.
6001 if (rt > env->fbq_type)
6004 power = power_of(i);
6005 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
6007 capacity = fix_small_capacity(env->sd, group);
6009 wl = weighted_cpuload(i);
6012 * When comparing with imbalance, use weighted_cpuload()
6013 * which is not scaled with the cpu power.
6015 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
6019 * For the load comparisons with the other cpu's, consider
6020 * the weighted_cpuload() scaled with the cpu power, so that
6021 * the load can be moved away from the cpu that is potentially
6022 * running at a lower capacity.
6024 * Thus we're looking for max(wl_i / power_i), crosswise
6025 * multiplication to rid ourselves of the division works out
6026 * to: wl_i * power_j > wl_j * power_i; where j is our
6029 if (wl * busiest_power > busiest_load * power) {
6031 busiest_power = power;
6040 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6041 * so long as it is large enough.
6043 #define MAX_PINNED_INTERVAL 512
6045 /* Working cpumask for load_balance and load_balance_newidle. */
6046 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6048 static int need_active_balance(struct lb_env *env)
6050 struct sched_domain *sd = env->sd;
6052 if (env->idle == CPU_NEWLY_IDLE) {
6055 * ASYM_PACKING needs to force migrate tasks from busy but
6056 * higher numbered CPUs in order to pack all tasks in the
6057 * lowest numbered CPUs.
6059 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6063 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6066 static int active_load_balance_cpu_stop(void *data);
6068 static int should_we_balance(struct lb_env *env)
6070 struct sched_group *sg = env->sd->groups;
6071 struct cpumask *sg_cpus, *sg_mask;
6072 int cpu, balance_cpu = -1;
6075 * In the newly idle case, we will allow all the cpu's
6076 * to do the newly idle load balance.
6078 if (env->idle == CPU_NEWLY_IDLE)
6081 sg_cpus = sched_group_cpus(sg);
6082 sg_mask = sched_group_mask(sg);
6083 /* Try to find first idle cpu */
6084 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6085 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6092 if (balance_cpu == -1)
6093 balance_cpu = group_balance_cpu(sg);
6096 * First idle cpu or the first cpu(busiest) in this sched group
6097 * is eligible for doing load balancing at this and above domains.
6099 return balance_cpu == env->dst_cpu;
6103 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6104 * tasks if there is an imbalance.
6106 static int load_balance(int this_cpu, struct rq *this_rq,
6107 struct sched_domain *sd, enum cpu_idle_type idle,
6108 int *continue_balancing)
6110 int ld_moved, cur_ld_moved, active_balance = 0;
6111 struct sched_domain *sd_parent = sd->parent;
6112 struct sched_group *group;
6114 unsigned long flags;
6115 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6117 struct lb_env env = {
6119 .dst_cpu = this_cpu,
6121 .dst_grpmask = sched_group_cpus(sd->groups),
6123 .loop_break = sched_nr_migrate_break,
6129 * For NEWLY_IDLE load_balancing, we don't need to consider
6130 * other cpus in our group
6132 if (idle == CPU_NEWLY_IDLE)
6133 env.dst_grpmask = NULL;
6135 cpumask_copy(cpus, cpu_active_mask);
6137 schedstat_inc(sd, lb_count[idle]);
6140 if (!should_we_balance(&env)) {
6141 *continue_balancing = 0;
6145 group = find_busiest_group(&env);
6147 schedstat_inc(sd, lb_nobusyg[idle]);
6151 busiest = find_busiest_queue(&env, group);
6153 schedstat_inc(sd, lb_nobusyq[idle]);
6157 BUG_ON(busiest == env.dst_rq);
6159 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6162 if (busiest->nr_running > 1) {
6164 * Attempt to move tasks. If find_busiest_group has found
6165 * an imbalance but busiest->nr_running <= 1, the group is
6166 * still unbalanced. ld_moved simply stays zero, so it is
6167 * correctly treated as an imbalance.
6169 env.flags |= LBF_ALL_PINNED;
6170 env.src_cpu = busiest->cpu;
6171 env.src_rq = busiest;
6172 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6175 local_irq_save(flags);
6176 double_rq_lock(env.dst_rq, busiest);
6179 * cur_ld_moved - load moved in current iteration
6180 * ld_moved - cumulative load moved across iterations
6182 cur_ld_moved = move_tasks(&env);
6183 ld_moved += cur_ld_moved;
6184 double_rq_unlock(env.dst_rq, busiest);
6185 local_irq_restore(flags);
6188 * some other cpu did the load balance for us.
6190 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6191 resched_cpu(env.dst_cpu);
6193 if (env.flags & LBF_NEED_BREAK) {
6194 env.flags &= ~LBF_NEED_BREAK;
6199 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6200 * us and move them to an alternate dst_cpu in our sched_group
6201 * where they can run. The upper limit on how many times we
6202 * iterate on same src_cpu is dependent on number of cpus in our
6205 * This changes load balance semantics a bit on who can move
6206 * load to a given_cpu. In addition to the given_cpu itself
6207 * (or a ilb_cpu acting on its behalf where given_cpu is
6208 * nohz-idle), we now have balance_cpu in a position to move
6209 * load to given_cpu. In rare situations, this may cause
6210 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6211 * _independently_ and at _same_ time to move some load to
6212 * given_cpu) causing exceess load to be moved to given_cpu.
6213 * This however should not happen so much in practice and
6214 * moreover subsequent load balance cycles should correct the
6215 * excess load moved.
6217 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6219 /* Prevent to re-select dst_cpu via env's cpus */
6220 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6222 env.dst_rq = cpu_rq(env.new_dst_cpu);
6223 env.dst_cpu = env.new_dst_cpu;
6224 env.flags &= ~LBF_DST_PINNED;
6226 env.loop_break = sched_nr_migrate_break;
6229 * Go back to "more_balance" rather than "redo" since we
6230 * need to continue with same src_cpu.
6236 * We failed to reach balance because of affinity.
6239 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6241 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6242 *group_imbalance = 1;
6243 } else if (*group_imbalance)
6244 *group_imbalance = 0;
6247 /* All tasks on this runqueue were pinned by CPU affinity */
6248 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6249 cpumask_clear_cpu(cpu_of(busiest), cpus);
6250 if (!cpumask_empty(cpus)) {
6252 env.loop_break = sched_nr_migrate_break;
6260 schedstat_inc(sd, lb_failed[idle]);
6262 * Increment the failure counter only on periodic balance.
6263 * We do not want newidle balance, which can be very
6264 * frequent, pollute the failure counter causing
6265 * excessive cache_hot migrations and active balances.
6267 if (idle != CPU_NEWLY_IDLE)
6268 sd->nr_balance_failed++;
6270 if (need_active_balance(&env)) {
6271 raw_spin_lock_irqsave(&busiest->lock, flags);
6273 /* don't kick the active_load_balance_cpu_stop,
6274 * if the curr task on busiest cpu can't be
6277 if (!cpumask_test_cpu(this_cpu,
6278 tsk_cpus_allowed(busiest->curr))) {
6279 raw_spin_unlock_irqrestore(&busiest->lock,
6281 env.flags |= LBF_ALL_PINNED;
6282 goto out_one_pinned;
6286 * ->active_balance synchronizes accesses to
6287 * ->active_balance_work. Once set, it's cleared
6288 * only after active load balance is finished.
6290 if (!busiest->active_balance) {
6291 busiest->active_balance = 1;
6292 busiest->push_cpu = this_cpu;
6295 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6297 if (active_balance) {
6298 stop_one_cpu_nowait(cpu_of(busiest),
6299 active_load_balance_cpu_stop, busiest,
6300 &busiest->active_balance_work);
6304 * We've kicked active balancing, reset the failure
6307 sd->nr_balance_failed = sd->cache_nice_tries+1;
6310 sd->nr_balance_failed = 0;
6312 if (likely(!active_balance)) {
6313 /* We were unbalanced, so reset the balancing interval */
6314 sd->balance_interval = sd->min_interval;
6317 * If we've begun active balancing, start to back off. This
6318 * case may not be covered by the all_pinned logic if there
6319 * is only 1 task on the busy runqueue (because we don't call
6322 if (sd->balance_interval < sd->max_interval)
6323 sd->balance_interval *= 2;
6329 schedstat_inc(sd, lb_balanced[idle]);
6331 sd->nr_balance_failed = 0;
6334 /* tune up the balancing interval */
6335 if (((env.flags & LBF_ALL_PINNED) &&
6336 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6337 (sd->balance_interval < sd->max_interval))
6338 sd->balance_interval *= 2;
6346 * idle_balance is called by schedule() if this_cpu is about to become
6347 * idle. Attempts to pull tasks from other CPUs.
6349 void idle_balance(int this_cpu, struct rq *this_rq)
6351 struct sched_domain *sd;
6352 int pulled_task = 0;
6353 unsigned long next_balance = jiffies + HZ;
6356 this_rq->idle_stamp = rq_clock(this_rq);
6358 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6362 * Drop the rq->lock, but keep IRQ/preempt disabled.
6364 raw_spin_unlock(&this_rq->lock);
6366 update_blocked_averages(this_cpu);
6368 for_each_domain(this_cpu, sd) {
6369 unsigned long interval;
6370 int continue_balancing = 1;
6371 u64 t0, domain_cost;
6373 if (!(sd->flags & SD_LOAD_BALANCE))
6376 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6379 if (sd->flags & SD_BALANCE_NEWIDLE) {
6380 t0 = sched_clock_cpu(this_cpu);
6382 /* If we've pulled tasks over stop searching: */
6383 pulled_task = load_balance(this_cpu, this_rq,
6385 &continue_balancing);
6387 domain_cost = sched_clock_cpu(this_cpu) - t0;
6388 if (domain_cost > sd->max_newidle_lb_cost)
6389 sd->max_newidle_lb_cost = domain_cost;
6391 curr_cost += domain_cost;
6394 interval = msecs_to_jiffies(sd->balance_interval);
6395 if (time_after(next_balance, sd->last_balance + interval))
6396 next_balance = sd->last_balance + interval;
6398 this_rq->idle_stamp = 0;
6404 raw_spin_lock(&this_rq->lock);
6406 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6408 * We are going idle. next_balance may be set based on
6409 * a busy processor. So reset next_balance.
6411 this_rq->next_balance = next_balance;
6414 if (curr_cost > this_rq->max_idle_balance_cost)
6415 this_rq->max_idle_balance_cost = curr_cost;
6419 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6420 * running tasks off the busiest CPU onto idle CPUs. It requires at
6421 * least 1 task to be running on each physical CPU where possible, and
6422 * avoids physical / logical imbalances.
6424 static int active_load_balance_cpu_stop(void *data)
6426 struct rq *busiest_rq = data;
6427 int busiest_cpu = cpu_of(busiest_rq);
6428 int target_cpu = busiest_rq->push_cpu;
6429 struct rq *target_rq = cpu_rq(target_cpu);
6430 struct sched_domain *sd;
6432 raw_spin_lock_irq(&busiest_rq->lock);
6434 /* make sure the requested cpu hasn't gone down in the meantime */
6435 if (unlikely(busiest_cpu != smp_processor_id() ||
6436 !busiest_rq->active_balance))
6439 /* Is there any task to move? */
6440 if (busiest_rq->nr_running <= 1)
6444 * This condition is "impossible", if it occurs
6445 * we need to fix it. Originally reported by
6446 * Bjorn Helgaas on a 128-cpu setup.
6448 BUG_ON(busiest_rq == target_rq);
6450 /* move a task from busiest_rq to target_rq */
6451 double_lock_balance(busiest_rq, target_rq);
6453 /* Search for an sd spanning us and the target CPU. */
6455 for_each_domain(target_cpu, sd) {
6456 if ((sd->flags & SD_LOAD_BALANCE) &&
6457 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6462 struct lb_env env = {
6464 .dst_cpu = target_cpu,
6465 .dst_rq = target_rq,
6466 .src_cpu = busiest_rq->cpu,
6467 .src_rq = busiest_rq,
6471 schedstat_inc(sd, alb_count);
6473 if (move_one_task(&env))
6474 schedstat_inc(sd, alb_pushed);
6476 schedstat_inc(sd, alb_failed);
6479 double_unlock_balance(busiest_rq, target_rq);
6481 busiest_rq->active_balance = 0;
6482 raw_spin_unlock_irq(&busiest_rq->lock);
6486 #ifdef CONFIG_NO_HZ_COMMON
6488 * idle load balancing details
6489 * - When one of the busy CPUs notice that there may be an idle rebalancing
6490 * needed, they will kick the idle load balancer, which then does idle
6491 * load balancing for all the idle CPUs.
6494 cpumask_var_t idle_cpus_mask;
6496 unsigned long next_balance; /* in jiffy units */
6497 } nohz ____cacheline_aligned;
6499 static inline int find_new_ilb(int call_cpu)
6501 int ilb = cpumask_first(nohz.idle_cpus_mask);
6503 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6510 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6511 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6512 * CPU (if there is one).
6514 static void nohz_balancer_kick(int cpu)
6518 nohz.next_balance++;
6520 ilb_cpu = find_new_ilb(cpu);
6522 if (ilb_cpu >= nr_cpu_ids)
6525 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6528 * Use smp_send_reschedule() instead of resched_cpu().
6529 * This way we generate a sched IPI on the target cpu which
6530 * is idle. And the softirq performing nohz idle load balance
6531 * will be run before returning from the IPI.
6533 smp_send_reschedule(ilb_cpu);
6537 static inline void nohz_balance_exit_idle(int cpu)
6539 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6540 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6541 atomic_dec(&nohz.nr_cpus);
6542 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6546 static inline void set_cpu_sd_state_busy(void)
6548 struct sched_domain *sd;
6549 int cpu = smp_processor_id();
6552 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6554 if (!sd || !sd->nohz_idle)
6558 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6563 void set_cpu_sd_state_idle(void)
6565 struct sched_domain *sd;
6566 int cpu = smp_processor_id();
6569 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6571 if (!sd || sd->nohz_idle)
6575 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6581 * This routine will record that the cpu is going idle with tick stopped.
6582 * This info will be used in performing idle load balancing in the future.
6584 void nohz_balance_enter_idle(int cpu)
6587 * If this cpu is going down, then nothing needs to be done.
6589 if (!cpu_active(cpu))
6592 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6595 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6596 atomic_inc(&nohz.nr_cpus);
6597 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6600 static int sched_ilb_notifier(struct notifier_block *nfb,
6601 unsigned long action, void *hcpu)
6603 switch (action & ~CPU_TASKS_FROZEN) {
6605 nohz_balance_exit_idle(smp_processor_id());
6613 static DEFINE_SPINLOCK(balancing);
6616 * Scale the max load_balance interval with the number of CPUs in the system.
6617 * This trades load-balance latency on larger machines for less cross talk.
6619 void update_max_interval(void)
6621 max_load_balance_interval = HZ*num_online_cpus()/10;
6625 * It checks each scheduling domain to see if it is due to be balanced,
6626 * and initiates a balancing operation if so.
6628 * Balancing parameters are set up in init_sched_domains.
6630 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6632 int continue_balancing = 1;
6633 struct rq *rq = cpu_rq(cpu);
6634 unsigned long interval;
6635 struct sched_domain *sd;
6636 /* Earliest time when we have to do rebalance again */
6637 unsigned long next_balance = jiffies + 60*HZ;
6638 int update_next_balance = 0;
6639 int need_serialize, need_decay = 0;
6642 update_blocked_averages(cpu);
6645 for_each_domain(cpu, sd) {
6647 * Decay the newidle max times here because this is a regular
6648 * visit to all the domains. Decay ~1% per second.
6650 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6651 sd->max_newidle_lb_cost =
6652 (sd->max_newidle_lb_cost * 253) / 256;
6653 sd->next_decay_max_lb_cost = jiffies + HZ;
6656 max_cost += sd->max_newidle_lb_cost;
6658 if (!(sd->flags & SD_LOAD_BALANCE))
6662 * Stop the load balance at this level. There is another
6663 * CPU in our sched group which is doing load balancing more
6666 if (!continue_balancing) {
6672 interval = sd->balance_interval;
6673 if (idle != CPU_IDLE)
6674 interval *= sd->busy_factor;
6676 /* scale ms to jiffies */
6677 interval = msecs_to_jiffies(interval);
6678 interval = clamp(interval, 1UL, max_load_balance_interval);
6680 need_serialize = sd->flags & SD_SERIALIZE;
6682 if (need_serialize) {
6683 if (!spin_trylock(&balancing))
6687 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6688 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6690 * The LBF_DST_PINNED logic could have changed
6691 * env->dst_cpu, so we can't know our idle
6692 * state even if we migrated tasks. Update it.
6694 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6696 sd->last_balance = jiffies;
6699 spin_unlock(&balancing);
6701 if (time_after(next_balance, sd->last_balance + interval)) {
6702 next_balance = sd->last_balance + interval;
6703 update_next_balance = 1;
6708 * Ensure the rq-wide value also decays but keep it at a
6709 * reasonable floor to avoid funnies with rq->avg_idle.
6711 rq->max_idle_balance_cost =
6712 max((u64)sysctl_sched_migration_cost, max_cost);
6717 * next_balance will be updated only when there is a need.
6718 * When the cpu is attached to null domain for ex, it will not be
6721 if (likely(update_next_balance))
6722 rq->next_balance = next_balance;
6725 #ifdef CONFIG_NO_HZ_COMMON
6727 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6728 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6730 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6732 struct rq *this_rq = cpu_rq(this_cpu);
6736 if (idle != CPU_IDLE ||
6737 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6740 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6741 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6745 * If this cpu gets work to do, stop the load balancing
6746 * work being done for other cpus. Next load
6747 * balancing owner will pick it up.
6752 rq = cpu_rq(balance_cpu);
6754 raw_spin_lock_irq(&rq->lock);
6755 update_rq_clock(rq);
6756 update_idle_cpu_load(rq);
6757 raw_spin_unlock_irq(&rq->lock);
6759 rebalance_domains(balance_cpu, CPU_IDLE);
6761 if (time_after(this_rq->next_balance, rq->next_balance))
6762 this_rq->next_balance = rq->next_balance;
6764 nohz.next_balance = this_rq->next_balance;
6766 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6770 * Current heuristic for kicking the idle load balancer in the presence
6771 * of an idle cpu is the system.
6772 * - This rq has more than one task.
6773 * - At any scheduler domain level, this cpu's scheduler group has multiple
6774 * busy cpu's exceeding the group's power.
6775 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6776 * domain span are idle.
6778 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6780 unsigned long now = jiffies;
6781 struct sched_domain *sd;
6782 struct sched_group_power *sgp;
6785 if (unlikely(idle_cpu(cpu)))
6789 * We may be recently in ticked or tickless idle mode. At the first
6790 * busy tick after returning from idle, we will update the busy stats.
6792 set_cpu_sd_state_busy();
6793 nohz_balance_exit_idle(cpu);
6796 * None are in tickless mode and hence no need for NOHZ idle load
6799 if (likely(!atomic_read(&nohz.nr_cpus)))
6802 if (time_before(now, nohz.next_balance))
6805 if (rq->nr_running >= 2)
6809 sd = rcu_dereference(per_cpu(sd_busy, cpu));
6812 sgp = sd->groups->sgp;
6813 nr_busy = atomic_read(&sgp->nr_busy_cpus);
6816 goto need_kick_unlock;
6819 sd = rcu_dereference(per_cpu(sd_asym, cpu));
6821 if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
6822 sched_domain_span(sd)) < cpu))
6823 goto need_kick_unlock;
6834 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6838 * run_rebalance_domains is triggered when needed from the scheduler tick.
6839 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6841 static void run_rebalance_domains(struct softirq_action *h)
6843 int this_cpu = smp_processor_id();
6844 struct rq *this_rq = cpu_rq(this_cpu);
6845 enum cpu_idle_type idle = this_rq->idle_balance ?
6846 CPU_IDLE : CPU_NOT_IDLE;
6848 rebalance_domains(this_cpu, idle);
6851 * If this cpu has a pending nohz_balance_kick, then do the
6852 * balancing on behalf of the other idle cpus whose ticks are
6855 nohz_idle_balance(this_cpu, idle);
6858 static inline int on_null_domain(int cpu)
6860 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6864 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6866 void trigger_load_balance(struct rq *rq, int cpu)
6868 /* Don't need to rebalance while attached to NULL domain */
6869 if (time_after_eq(jiffies, rq->next_balance) &&
6870 likely(!on_null_domain(cpu)))
6871 raise_softirq(SCHED_SOFTIRQ);
6872 #ifdef CONFIG_NO_HZ_COMMON
6873 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6874 nohz_balancer_kick(cpu);
6878 static void rq_online_fair(struct rq *rq)
6883 static void rq_offline_fair(struct rq *rq)
6887 /* Ensure any throttled groups are reachable by pick_next_task */
6888 unthrottle_offline_cfs_rqs(rq);
6891 #endif /* CONFIG_SMP */
6894 * scheduler tick hitting a task of our scheduling class:
6896 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6898 struct cfs_rq *cfs_rq;
6899 struct sched_entity *se = &curr->se;
6901 for_each_sched_entity(se) {
6902 cfs_rq = cfs_rq_of(se);
6903 entity_tick(cfs_rq, se, queued);
6906 if (numabalancing_enabled)
6907 task_tick_numa(rq, curr);
6909 update_rq_runnable_avg(rq, 1);
6913 * called on fork with the child task as argument from the parent's context
6914 * - child not yet on the tasklist
6915 * - preemption disabled
6917 static void task_fork_fair(struct task_struct *p)
6919 struct cfs_rq *cfs_rq;
6920 struct sched_entity *se = &p->se, *curr;
6921 int this_cpu = smp_processor_id();
6922 struct rq *rq = this_rq();
6923 unsigned long flags;
6925 raw_spin_lock_irqsave(&rq->lock, flags);
6927 update_rq_clock(rq);
6929 cfs_rq = task_cfs_rq(current);
6930 curr = cfs_rq->curr;
6933 * Not only the cpu but also the task_group of the parent might have
6934 * been changed after parent->se.parent,cfs_rq were copied to
6935 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6936 * of child point to valid ones.
6939 __set_task_cpu(p, this_cpu);
6942 update_curr(cfs_rq);
6945 se->vruntime = curr->vruntime;
6946 place_entity(cfs_rq, se, 1);
6948 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6950 * Upon rescheduling, sched_class::put_prev_task() will place
6951 * 'current' within the tree based on its new key value.
6953 swap(curr->vruntime, se->vruntime);
6954 resched_task(rq->curr);
6957 se->vruntime -= cfs_rq->min_vruntime;
6959 raw_spin_unlock_irqrestore(&rq->lock, flags);
6963 * Priority of the task has changed. Check to see if we preempt
6967 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6973 * Reschedule if we are currently running on this runqueue and
6974 * our priority decreased, or if we are not currently running on
6975 * this runqueue and our priority is higher than the current's
6977 if (rq->curr == p) {
6978 if (p->prio > oldprio)
6979 resched_task(rq->curr);
6981 check_preempt_curr(rq, p, 0);
6984 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6986 struct sched_entity *se = &p->se;
6987 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6990 * Ensure the task's vruntime is normalized, so that when its
6991 * switched back to the fair class the enqueue_entity(.flags=0) will
6992 * do the right thing.
6994 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6995 * have normalized the vruntime, if it was !on_rq, then only when
6996 * the task is sleeping will it still have non-normalized vruntime.
6998 if (!se->on_rq && p->state != TASK_RUNNING) {
7000 * Fix up our vruntime so that the current sleep doesn't
7001 * cause 'unlimited' sleep bonus.
7003 place_entity(cfs_rq, se, 0);
7004 se->vruntime -= cfs_rq->min_vruntime;
7009 * Remove our load from contribution when we leave sched_fair
7010 * and ensure we don't carry in an old decay_count if we
7013 if (se->avg.decay_count) {
7014 __synchronize_entity_decay(se);
7015 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
7021 * We switched to the sched_fair class.
7023 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7029 * We were most likely switched from sched_rt, so
7030 * kick off the schedule if running, otherwise just see
7031 * if we can still preempt the current task.
7034 resched_task(rq->curr);
7036 check_preempt_curr(rq, p, 0);
7039 /* Account for a task changing its policy or group.
7041 * This routine is mostly called to set cfs_rq->curr field when a task
7042 * migrates between groups/classes.
7044 static void set_curr_task_fair(struct rq *rq)
7046 struct sched_entity *se = &rq->curr->se;
7048 for_each_sched_entity(se) {
7049 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7051 set_next_entity(cfs_rq, se);
7052 /* ensure bandwidth has been allocated on our new cfs_rq */
7053 account_cfs_rq_runtime(cfs_rq, 0);
7057 void init_cfs_rq(struct cfs_rq *cfs_rq)
7059 cfs_rq->tasks_timeline = RB_ROOT;
7060 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7061 #ifndef CONFIG_64BIT
7062 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7065 atomic64_set(&cfs_rq->decay_counter, 1);
7066 atomic_long_set(&cfs_rq->removed_load, 0);
7070 #ifdef CONFIG_FAIR_GROUP_SCHED
7071 static void task_move_group_fair(struct task_struct *p, int on_rq)
7073 struct cfs_rq *cfs_rq;
7075 * If the task was not on the rq at the time of this cgroup movement
7076 * it must have been asleep, sleeping tasks keep their ->vruntime
7077 * absolute on their old rq until wakeup (needed for the fair sleeper
7078 * bonus in place_entity()).
7080 * If it was on the rq, we've just 'preempted' it, which does convert
7081 * ->vruntime to a relative base.
7083 * Make sure both cases convert their relative position when migrating
7084 * to another cgroup's rq. This does somewhat interfere with the
7085 * fair sleeper stuff for the first placement, but who cares.
7088 * When !on_rq, vruntime of the task has usually NOT been normalized.
7089 * But there are some cases where it has already been normalized:
7091 * - Moving a forked child which is waiting for being woken up by
7092 * wake_up_new_task().
7093 * - Moving a task which has been woken up by try_to_wake_up() and
7094 * waiting for actually being woken up by sched_ttwu_pending().
7096 * To prevent boost or penalty in the new cfs_rq caused by delta
7097 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7099 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7103 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7104 set_task_rq(p, task_cpu(p));
7106 cfs_rq = cfs_rq_of(&p->se);
7107 p->se.vruntime += cfs_rq->min_vruntime;
7110 * migrate_task_rq_fair() will have removed our previous
7111 * contribution, but we must synchronize for ongoing future
7114 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7115 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7120 void free_fair_sched_group(struct task_group *tg)
7124 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7126 for_each_possible_cpu(i) {
7128 kfree(tg->cfs_rq[i]);
7137 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7139 struct cfs_rq *cfs_rq;
7140 struct sched_entity *se;
7143 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7146 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7150 tg->shares = NICE_0_LOAD;
7152 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7154 for_each_possible_cpu(i) {
7155 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7156 GFP_KERNEL, cpu_to_node(i));
7160 se = kzalloc_node(sizeof(struct sched_entity),
7161 GFP_KERNEL, cpu_to_node(i));
7165 init_cfs_rq(cfs_rq);
7166 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7177 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7179 struct rq *rq = cpu_rq(cpu);
7180 unsigned long flags;
7183 * Only empty task groups can be destroyed; so we can speculatively
7184 * check on_list without danger of it being re-added.
7186 if (!tg->cfs_rq[cpu]->on_list)
7189 raw_spin_lock_irqsave(&rq->lock, flags);
7190 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7191 raw_spin_unlock_irqrestore(&rq->lock, flags);
7194 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7195 struct sched_entity *se, int cpu,
7196 struct sched_entity *parent)
7198 struct rq *rq = cpu_rq(cpu);
7202 init_cfs_rq_runtime(cfs_rq);
7204 tg->cfs_rq[cpu] = cfs_rq;
7207 /* se could be NULL for root_task_group */
7212 se->cfs_rq = &rq->cfs;
7214 se->cfs_rq = parent->my_q;
7217 /* guarantee group entities always have weight */
7218 update_load_set(&se->load, NICE_0_LOAD);
7219 se->parent = parent;
7222 static DEFINE_MUTEX(shares_mutex);
7224 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7227 unsigned long flags;
7230 * We can't change the weight of the root cgroup.
7235 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7237 mutex_lock(&shares_mutex);
7238 if (tg->shares == shares)
7241 tg->shares = shares;
7242 for_each_possible_cpu(i) {
7243 struct rq *rq = cpu_rq(i);
7244 struct sched_entity *se;
7247 /* Propagate contribution to hierarchy */
7248 raw_spin_lock_irqsave(&rq->lock, flags);
7250 /* Possible calls to update_curr() need rq clock */
7251 update_rq_clock(rq);
7252 for_each_sched_entity(se)
7253 update_cfs_shares(group_cfs_rq(se));
7254 raw_spin_unlock_irqrestore(&rq->lock, flags);
7258 mutex_unlock(&shares_mutex);
7261 #else /* CONFIG_FAIR_GROUP_SCHED */
7263 void free_fair_sched_group(struct task_group *tg) { }
7265 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7270 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7272 #endif /* CONFIG_FAIR_GROUP_SCHED */
7275 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7277 struct sched_entity *se = &task->se;
7278 unsigned int rr_interval = 0;
7281 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7284 if (rq->cfs.load.weight)
7285 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7291 * All the scheduling class methods:
7293 const struct sched_class fair_sched_class = {
7294 .next = &idle_sched_class,
7295 .enqueue_task = enqueue_task_fair,
7296 .dequeue_task = dequeue_task_fair,
7297 .yield_task = yield_task_fair,
7298 .yield_to_task = yield_to_task_fair,
7300 .check_preempt_curr = check_preempt_wakeup,
7302 .pick_next_task = pick_next_task_fair,
7303 .put_prev_task = put_prev_task_fair,
7306 .select_task_rq = select_task_rq_fair,
7307 .migrate_task_rq = migrate_task_rq_fair,
7309 .rq_online = rq_online_fair,
7310 .rq_offline = rq_offline_fair,
7312 .task_waking = task_waking_fair,
7315 .set_curr_task = set_curr_task_fair,
7316 .task_tick = task_tick_fair,
7317 .task_fork = task_fork_fair,
7319 .prio_changed = prio_changed_fair,
7320 .switched_from = switched_from_fair,
7321 .switched_to = switched_to_fair,
7323 .get_rr_interval = get_rr_interval_fair,
7325 #ifdef CONFIG_FAIR_GROUP_SCHED
7326 .task_move_group = task_move_group_fair,
7330 #ifdef CONFIG_SCHED_DEBUG
7331 void print_cfs_stats(struct seq_file *m, int cpu)
7333 struct cfs_rq *cfs_rq;
7336 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7337 print_cfs_rq(m, cpu, cfs_rq);
7342 __init void init_sched_fair_class(void)
7345 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7347 #ifdef CONFIG_NO_HZ_COMMON
7348 nohz.next_balance = jiffies;
7349 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7350 cpu_notifier(sched_ilb_notifier, 0);