2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
29 #include <linux/mempolicy.h>
30 #include <linux/migrate.h>
31 #include <linux/task_work.h>
33 #include <trace/events/sched.h>
38 * Targeted preemption latency for CPU-bound tasks:
39 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
41 * NOTE: this latency value is not the same as the concept of
42 * 'timeslice length' - timeslices in CFS are of variable length
43 * and have no persistent notion like in traditional, time-slice
44 * based scheduling concepts.
46 * (to see the precise effective timeslice length of your workload,
47 * run vmstat and monitor the context-switches (cs) field)
49 unsigned int sysctl_sched_latency = 6000000ULL;
50 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
53 * The initial- and re-scaling of tunables is configurable
54 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
57 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
58 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
59 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
61 enum sched_tunable_scaling sysctl_sched_tunable_scaling
62 = SCHED_TUNABLESCALING_LOG;
65 * Minimal preemption granularity for CPU-bound tasks:
66 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 unsigned int sysctl_sched_min_granularity = 750000ULL;
69 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
72 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
74 static unsigned int sched_nr_latency = 8;
77 * After fork, child runs first. If set to 0 (default) then
78 * parent will (try to) run first.
80 unsigned int sysctl_sched_child_runs_first __read_mostly;
83 * SCHED_OTHER wake-up granularity.
84 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
86 * This option delays the preemption effects of decoupled workloads
87 * and reduces their over-scheduling. Synchronous workloads will still
88 * have immediate wakeup/sleep latencies.
90 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
91 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
93 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
96 * The exponential sliding window over which load is averaged for shares
100 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 #ifdef CONFIG_CFS_BANDWIDTH
104 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
105 * each time a cfs_rq requests quota.
107 * Note: in the case that the slice exceeds the runtime remaining (either due
108 * to consumption or the quota being specified to be smaller than the slice)
109 * we will always only issue the remaining available time.
111 * default: 5 msec, units: microseconds
113 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
116 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128 static inline void update_load_set(struct load_weight *lw, unsigned long w)
135 * Increase the granularity value when there are more CPUs,
136 * because with more CPUs the 'effective latency' as visible
137 * to users decreases. But the relationship is not linear,
138 * so pick a second-best guess by going with the log2 of the
141 * This idea comes from the SD scheduler of Con Kolivas:
143 static int get_update_sysctl_factor(void)
145 unsigned int cpus = min_t(int, num_online_cpus(), 8);
148 switch (sysctl_sched_tunable_scaling) {
149 case SCHED_TUNABLESCALING_NONE:
152 case SCHED_TUNABLESCALING_LINEAR:
155 case SCHED_TUNABLESCALING_LOG:
157 factor = 1 + ilog2(cpus);
164 static void update_sysctl(void)
166 unsigned int factor = get_update_sysctl_factor();
168 #define SET_SYSCTL(name) \
169 (sysctl_##name = (factor) * normalized_sysctl_##name)
170 SET_SYSCTL(sched_min_granularity);
171 SET_SYSCTL(sched_latency);
172 SET_SYSCTL(sched_wakeup_granularity);
176 void sched_init_granularity(void)
181 #if BITS_PER_LONG == 32
182 # define WMULT_CONST (~0UL)
184 # define WMULT_CONST (1UL << 32)
187 #define WMULT_SHIFT 32
190 * Shift right and round:
192 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
195 * delta *= weight / lw
198 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
199 struct load_weight *lw)
204 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
205 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
206 * 2^SCHED_LOAD_RESOLUTION.
208 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
209 tmp = (u64)delta_exec * scale_load_down(weight);
211 tmp = (u64)delta_exec;
213 if (!lw->inv_weight) {
214 unsigned long w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * Check whether we'd overflow the 64-bit multiplication:
227 if (unlikely(tmp > WMULT_CONST))
228 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
231 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
233 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
237 const struct sched_class fair_sched_class;
239 /**************************************************************
240 * CFS operations on generic schedulable entities:
243 #ifdef CONFIG_FAIR_GROUP_SCHED
245 /* cpu runqueue to which this cfs_rq is attached */
246 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
251 /* An entity is a task if it doesn't "own" a runqueue */
252 #define entity_is_task(se) (!se->my_q)
254 static inline struct task_struct *task_of(struct sched_entity *se)
256 #ifdef CONFIG_SCHED_DEBUG
257 WARN_ON_ONCE(!entity_is_task(se));
259 return container_of(se, struct task_struct, se);
262 /* Walk up scheduling entities hierarchy */
263 #define for_each_sched_entity(se) \
264 for (; se; se = se->parent)
266 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
271 /* runqueue on which this entity is (to be) queued */
272 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
277 /* runqueue "owned" by this group */
278 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
283 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 if (!cfs_rq->on_list) {
290 * Ensure we either appear before our parent (if already
291 * enqueued) or force our parent to appear after us when it is
292 * enqueued. The fact that we always enqueue bottom-up
293 * reduces this to two cases.
295 if (cfs_rq->tg->parent &&
296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298 &rq_of(cfs_rq)->leaf_cfs_rq_list);
300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301 &rq_of(cfs_rq)->leaf_cfs_rq_list);
305 /* We should have no load, but we need to update last_decay. */
306 update_cfs_rq_blocked_load(cfs_rq, 0);
310 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
312 if (cfs_rq->on_list) {
313 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
318 /* Iterate thr' all leaf cfs_rq's on a runqueue */
319 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
320 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
322 /* Do the two (enqueued) entities belong to the same group ? */
324 is_same_group(struct sched_entity *se, struct sched_entity *pse)
326 if (se->cfs_rq == pse->cfs_rq)
332 static inline struct sched_entity *parent_entity(struct sched_entity *se)
337 /* return depth at which a sched entity is present in the hierarchy */
338 static inline int depth_se(struct sched_entity *se)
342 for_each_sched_entity(se)
349 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
351 int se_depth, pse_depth;
354 * preemption test can be made between sibling entities who are in the
355 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
356 * both tasks until we find their ancestors who are siblings of common
360 /* First walk up until both entities are at same depth */
361 se_depth = depth_se(*se);
362 pse_depth = depth_se(*pse);
364 while (se_depth > pse_depth) {
366 *se = parent_entity(*se);
369 while (pse_depth > se_depth) {
371 *pse = parent_entity(*pse);
374 while (!is_same_group(*se, *pse)) {
375 *se = parent_entity(*se);
376 *pse = parent_entity(*pse);
380 #else /* !CONFIG_FAIR_GROUP_SCHED */
382 static inline struct task_struct *task_of(struct sched_entity *se)
384 return container_of(se, struct task_struct, se);
387 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
389 return container_of(cfs_rq, struct rq, cfs);
392 #define entity_is_task(se) 1
394 #define for_each_sched_entity(se) \
395 for (; se; se = NULL)
397 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
399 return &task_rq(p)->cfs;
402 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
404 struct task_struct *p = task_of(se);
405 struct rq *rq = task_rq(p);
410 /* runqueue "owned" by this group */
411 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
416 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
420 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
424 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
425 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
428 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 static inline struct sched_entity *parent_entity(struct sched_entity *se)
439 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
443 #endif /* CONFIG_FAIR_GROUP_SCHED */
445 static __always_inline
446 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
448 /**************************************************************
449 * Scheduling class tree data structure manipulation methods:
452 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
454 s64 delta = (s64)(vruntime - max_vruntime);
456 max_vruntime = vruntime;
461 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
463 s64 delta = (s64)(vruntime - min_vruntime);
465 min_vruntime = vruntime;
470 static inline int entity_before(struct sched_entity *a,
471 struct sched_entity *b)
473 return (s64)(a->vruntime - b->vruntime) < 0;
476 static void update_min_vruntime(struct cfs_rq *cfs_rq)
478 u64 vruntime = cfs_rq->min_vruntime;
481 vruntime = cfs_rq->curr->vruntime;
483 if (cfs_rq->rb_leftmost) {
484 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
489 vruntime = se->vruntime;
491 vruntime = min_vruntime(vruntime, se->vruntime);
494 /* ensure we never gain time by being placed backwards. */
495 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
498 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
503 * Enqueue an entity into the rb-tree:
505 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
507 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
508 struct rb_node *parent = NULL;
509 struct sched_entity *entry;
513 * Find the right place in the rbtree:
517 entry = rb_entry(parent, struct sched_entity, run_node);
519 * We dont care about collisions. Nodes with
520 * the same key stay together.
522 if (entity_before(se, entry)) {
523 link = &parent->rb_left;
525 link = &parent->rb_right;
531 * Maintain a cache of leftmost tree entries (it is frequently
535 cfs_rq->rb_leftmost = &se->run_node;
537 rb_link_node(&se->run_node, parent, link);
538 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
541 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 if (cfs_rq->rb_leftmost == &se->run_node) {
544 struct rb_node *next_node;
546 next_node = rb_next(&se->run_node);
547 cfs_rq->rb_leftmost = next_node;
550 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
553 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
555 struct rb_node *left = cfs_rq->rb_leftmost;
560 return rb_entry(left, struct sched_entity, run_node);
563 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
565 struct rb_node *next = rb_next(&se->run_node);
570 return rb_entry(next, struct sched_entity, run_node);
573 #ifdef CONFIG_SCHED_DEBUG
574 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
576 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
581 return rb_entry(last, struct sched_entity, run_node);
584 /**************************************************************
585 * Scheduling class statistics methods:
588 int sched_proc_update_handler(struct ctl_table *table, int write,
589 void __user *buffer, size_t *lenp,
592 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
593 int factor = get_update_sysctl_factor();
598 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
599 sysctl_sched_min_granularity);
601 #define WRT_SYSCTL(name) \
602 (normalized_sysctl_##name = sysctl_##name / (factor))
603 WRT_SYSCTL(sched_min_granularity);
604 WRT_SYSCTL(sched_latency);
605 WRT_SYSCTL(sched_wakeup_granularity);
615 static inline unsigned long
616 calc_delta_fair(unsigned long delta, struct sched_entity *se)
618 if (unlikely(se->load.weight != NICE_0_LOAD))
619 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
625 * The idea is to set a period in which each task runs once.
627 * When there are too many tasks (sched_nr_latency) we have to stretch
628 * this period because otherwise the slices get too small.
630 * p = (nr <= nl) ? l : l*nr/nl
632 static u64 __sched_period(unsigned long nr_running)
634 u64 period = sysctl_sched_latency;
635 unsigned long nr_latency = sched_nr_latency;
637 if (unlikely(nr_running > nr_latency)) {
638 period = sysctl_sched_min_granularity;
639 period *= nr_running;
646 * We calculate the wall-time slice from the period by taking a part
647 * proportional to the weight.
651 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
655 for_each_sched_entity(se) {
656 struct load_weight *load;
657 struct load_weight lw;
659 cfs_rq = cfs_rq_of(se);
660 load = &cfs_rq->load;
662 if (unlikely(!se->on_rq)) {
665 update_load_add(&lw, se->load.weight);
668 slice = calc_delta_mine(slice, se->load.weight, load);
674 * We calculate the vruntime slice of a to-be-inserted task.
678 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 return calc_delta_fair(sched_slice(cfs_rq, se), se);
684 static unsigned long task_h_load(struct task_struct *p);
686 static inline void __update_task_entity_contrib(struct sched_entity *se);
688 /* Give new task start runnable values to heavy its load in infant time */
689 void init_task_runnable_average(struct task_struct *p)
693 p->se.avg.decay_count = 0;
694 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10;
695 p->se.avg.runnable_avg_sum = slice;
696 p->se.avg.runnable_avg_period = slice;
697 __update_task_entity_contrib(&p->se);
700 void init_task_runnable_average(struct task_struct *p)
706 * Update the current task's runtime statistics. Skip current tasks that
707 * are not in our scheduling class.
710 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
711 unsigned long delta_exec)
713 unsigned long delta_exec_weighted;
715 schedstat_set(curr->statistics.exec_max,
716 max((u64)delta_exec, curr->statistics.exec_max));
718 curr->sum_exec_runtime += delta_exec;
719 schedstat_add(cfs_rq, exec_clock, delta_exec);
720 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
722 curr->vruntime += delta_exec_weighted;
723 update_min_vruntime(cfs_rq);
726 static void update_curr(struct cfs_rq *cfs_rq)
728 struct sched_entity *curr = cfs_rq->curr;
729 u64 now = rq_clock_task(rq_of(cfs_rq));
730 unsigned long delta_exec;
736 * Get the amount of time the current task was running
737 * since the last time we changed load (this cannot
738 * overflow on 32 bits):
740 delta_exec = (unsigned long)(now - curr->exec_start);
744 __update_curr(cfs_rq, curr, delta_exec);
745 curr->exec_start = now;
747 if (entity_is_task(curr)) {
748 struct task_struct *curtask = task_of(curr);
750 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
751 cpuacct_charge(curtask, delta_exec);
752 account_group_exec_runtime(curtask, delta_exec);
755 account_cfs_rq_runtime(cfs_rq, delta_exec);
759 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
761 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq)));
765 * Task is being enqueued - update stats:
767 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
770 * Are we enqueueing a waiting task? (for current tasks
771 * a dequeue/enqueue event is a NOP)
773 if (se != cfs_rq->curr)
774 update_stats_wait_start(cfs_rq, se);
778 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
781 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start));
782 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
783 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
784 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
785 #ifdef CONFIG_SCHEDSTATS
786 if (entity_is_task(se)) {
787 trace_sched_stat_wait(task_of(se),
788 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start);
791 schedstat_set(se->statistics.wait_start, 0);
795 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 * Mark the end of the wait period if dequeueing a
801 if (se != cfs_rq->curr)
802 update_stats_wait_end(cfs_rq, se);
806 * We are picking a new current task - update its stats:
809 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
812 * We are starting a new run period:
814 se->exec_start = rq_clock_task(rq_of(cfs_rq));
817 /**************************************************
818 * Scheduling class queueing methods:
821 #ifdef CONFIG_NUMA_BALANCING
823 * Approximate time to scan a full NUMA task in ms. The task scan period is
824 * calculated based on the tasks virtual memory size and
825 * numa_balancing_scan_size.
827 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
828 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
830 /* Portion of address space to scan in MB */
831 unsigned int sysctl_numa_balancing_scan_size = 256;
833 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
834 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 * After skipping a page migration on a shared page, skip N more numa page
838 * migrations unconditionally. This reduces the number of NUMA migrations
839 * in shared memory workloads, and has the effect of pulling tasks towards
840 * where their memory lives, over pulling the memory towards the task.
842 unsigned int sysctl_numa_balancing_migrate_deferred = 16;
844 static unsigned int task_nr_scan_windows(struct task_struct *p)
846 unsigned long rss = 0;
847 unsigned long nr_scan_pages;
850 * Calculations based on RSS as non-present and empty pages are skipped
851 * by the PTE scanner and NUMA hinting faults should be trapped based
854 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
855 rss = get_mm_rss(p->mm);
859 rss = round_up(rss, nr_scan_pages);
860 return rss / nr_scan_pages;
863 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
864 #define MAX_SCAN_WINDOW 2560
866 static unsigned int task_scan_min(struct task_struct *p)
868 unsigned int scan, floor;
869 unsigned int windows = 1;
871 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
872 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
873 floor = 1000 / windows;
875 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
876 return max_t(unsigned int, floor, scan);
879 static unsigned int task_scan_max(struct task_struct *p)
881 unsigned int smin = task_scan_min(p);
884 /* Watch for min being lower than max due to floor calculations */
885 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
886 return max(smin, smax);
890 * Once a preferred node is selected the scheduler balancer will prefer moving
891 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
892 * scans. This will give the process the chance to accumulate more faults on
893 * the preferred node but still allow the scheduler to move the task again if
894 * the nodes CPUs are overloaded.
896 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
898 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
900 rq->nr_numa_running += (p->numa_preferred_nid != -1);
901 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
904 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
906 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
907 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
913 spinlock_t lock; /* nr_tasks, tasks */
916 struct list_head task_list;
919 unsigned long total_faults;
920 unsigned long faults[0];
923 pid_t task_numa_group_id(struct task_struct *p)
925 return p->numa_group ? p->numa_group->gid : 0;
928 static inline int task_faults_idx(int nid, int priv)
930 return 2 * nid + priv;
933 static inline unsigned long task_faults(struct task_struct *p, int nid)
938 return p->numa_faults[task_faults_idx(nid, 0)] +
939 p->numa_faults[task_faults_idx(nid, 1)];
942 static inline unsigned long group_faults(struct task_struct *p, int nid)
947 return p->numa_group->faults[2*nid] + p->numa_group->faults[2*nid+1];
951 * These return the fraction of accesses done by a particular task, or
952 * task group, on a particular numa node. The group weight is given a
953 * larger multiplier, in order to group tasks together that are almost
954 * evenly spread out between numa nodes.
956 static inline unsigned long task_weight(struct task_struct *p, int nid)
958 unsigned long total_faults;
963 total_faults = p->total_numa_faults;
968 return 1000 * task_faults(p, nid) / total_faults;
971 static inline unsigned long group_weight(struct task_struct *p, int nid)
973 if (!p->numa_group || !p->numa_group->total_faults)
976 return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
979 static unsigned long weighted_cpuload(const int cpu);
980 static unsigned long source_load(int cpu, int type);
981 static unsigned long target_load(int cpu, int type);
982 static unsigned long power_of(int cpu);
983 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
985 /* Cached statistics for all CPUs within a node */
987 unsigned long nr_running;
990 /* Total compute capacity of CPUs on a node */
993 /* Approximate capacity in terms of runnable tasks on a node */
994 unsigned long capacity;
999 * XXX borrowed from update_sg_lb_stats
1001 static void update_numa_stats(struct numa_stats *ns, int nid)
1005 memset(ns, 0, sizeof(*ns));
1006 for_each_cpu(cpu, cpumask_of_node(nid)) {
1007 struct rq *rq = cpu_rq(cpu);
1009 ns->nr_running += rq->nr_running;
1010 ns->load += weighted_cpuload(cpu);
1011 ns->power += power_of(cpu);
1014 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1015 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1016 ns->has_capacity = (ns->nr_running < ns->capacity);
1019 struct task_numa_env {
1020 struct task_struct *p;
1022 int src_cpu, src_nid;
1023 int dst_cpu, dst_nid;
1025 struct numa_stats src_stats, dst_stats;
1027 int imbalance_pct, idx;
1029 struct task_struct *best_task;
1034 static void task_numa_assign(struct task_numa_env *env,
1035 struct task_struct *p, long imp)
1038 put_task_struct(env->best_task);
1043 env->best_imp = imp;
1044 env->best_cpu = env->dst_cpu;
1048 * This checks if the overall compute and NUMA accesses of the system would
1049 * be improved if the source tasks was migrated to the target dst_cpu taking
1050 * into account that it might be best if task running on the dst_cpu should
1051 * be exchanged with the source task
1053 static void task_numa_compare(struct task_numa_env *env,
1054 long taskimp, long groupimp)
1056 struct rq *src_rq = cpu_rq(env->src_cpu);
1057 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1058 struct task_struct *cur;
1059 long dst_load, src_load;
1061 long imp = (groupimp > 0) ? groupimp : taskimp;
1064 cur = ACCESS_ONCE(dst_rq->curr);
1065 if (cur->pid == 0) /* idle */
1069 * "imp" is the fault differential for the source task between the
1070 * source and destination node. Calculate the total differential for
1071 * the source task and potential destination task. The more negative
1072 * the value is, the more rmeote accesses that would be expected to
1073 * be incurred if the tasks were swapped.
1076 /* Skip this swap candidate if cannot move to the source cpu */
1077 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1081 * If dst and source tasks are in the same NUMA group, or not
1082 * in any group then look only at task weights.
1084 if (cur->numa_group == env->p->numa_group) {
1085 imp = taskimp + task_weight(cur, env->src_nid) -
1086 task_weight(cur, env->dst_nid);
1088 * Add some hysteresis to prevent swapping the
1089 * tasks within a group over tiny differences.
1091 if (cur->numa_group)
1095 * Compare the group weights. If a task is all by
1096 * itself (not part of a group), use the task weight
1099 if (env->p->numa_group)
1104 if (cur->numa_group)
1105 imp += group_weight(cur, env->src_nid) -
1106 group_weight(cur, env->dst_nid);
1108 imp += task_weight(cur, env->src_nid) -
1109 task_weight(cur, env->dst_nid);
1113 if (imp < env->best_imp)
1117 /* Is there capacity at our destination? */
1118 if (env->src_stats.has_capacity &&
1119 !env->dst_stats.has_capacity)
1125 /* Balance doesn't matter much if we're running a task per cpu */
1126 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1130 * In the overloaded case, try and keep the load balanced.
1133 dst_load = env->dst_stats.load;
1134 src_load = env->src_stats.load;
1136 /* XXX missing power terms */
1137 load = task_h_load(env->p);
1142 load = task_h_load(cur);
1147 /* make src_load the smaller */
1148 if (dst_load < src_load)
1149 swap(dst_load, src_load);
1151 if (src_load * env->imbalance_pct < dst_load * 100)
1155 task_numa_assign(env, cur, imp);
1160 static void task_numa_find_cpu(struct task_numa_env *env,
1161 long taskimp, long groupimp)
1165 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1166 /* Skip this CPU if the source task cannot migrate */
1167 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1171 task_numa_compare(env, taskimp, groupimp);
1175 static int task_numa_migrate(struct task_struct *p)
1177 struct task_numa_env env = {
1180 .src_cpu = task_cpu(p),
1181 .src_nid = task_node(p),
1183 .imbalance_pct = 112,
1189 struct sched_domain *sd;
1190 unsigned long taskweight, groupweight;
1192 long taskimp, groupimp;
1195 * Pick the lowest SD_NUMA domain, as that would have the smallest
1196 * imbalance and would be the first to start moving tasks about.
1198 * And we want to avoid any moving of tasks about, as that would create
1199 * random movement of tasks -- counter the numa conditions we're trying
1203 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1204 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1207 taskweight = task_weight(p, env.src_nid);
1208 groupweight = group_weight(p, env.src_nid);
1209 update_numa_stats(&env.src_stats, env.src_nid);
1210 env.dst_nid = p->numa_preferred_nid;
1211 taskimp = task_weight(p, env.dst_nid) - taskweight;
1212 groupimp = group_weight(p, env.dst_nid) - groupweight;
1213 update_numa_stats(&env.dst_stats, env.dst_nid);
1215 /* If the preferred nid has capacity, try to use it. */
1216 if (env.dst_stats.has_capacity)
1217 task_numa_find_cpu(&env, taskimp, groupimp);
1219 /* No space available on the preferred nid. Look elsewhere. */
1220 if (env.best_cpu == -1) {
1221 for_each_online_node(nid) {
1222 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1225 /* Only consider nodes where both task and groups benefit */
1226 taskimp = task_weight(p, nid) - taskweight;
1227 groupimp = group_weight(p, nid) - groupweight;
1228 if (taskimp < 0 && groupimp < 0)
1232 update_numa_stats(&env.dst_stats, env.dst_nid);
1233 task_numa_find_cpu(&env, taskimp, groupimp);
1237 /* No better CPU than the current one was found. */
1238 if (env.best_cpu == -1)
1241 sched_setnuma(p, env.dst_nid);
1244 * Reset the scan period if the task is being rescheduled on an
1245 * alternative node to recheck if the tasks is now properly placed.
1247 p->numa_scan_period = task_scan_min(p);
1249 if (env.best_task == NULL) {
1250 int ret = migrate_task_to(p, env.best_cpu);
1254 ret = migrate_swap(p, env.best_task);
1255 put_task_struct(env.best_task);
1259 /* Attempt to migrate a task to a CPU on the preferred node. */
1260 static void numa_migrate_preferred(struct task_struct *p)
1262 /* This task has no NUMA fault statistics yet */
1263 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1266 /* Periodically retry migrating the task to the preferred node */
1267 p->numa_migrate_retry = jiffies + HZ;
1269 /* Success if task is already running on preferred CPU */
1270 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid)
1273 /* Otherwise, try migrate to a CPU on the preferred node */
1274 task_numa_migrate(p);
1278 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1279 * increments. The more local the fault statistics are, the higher the scan
1280 * period will be for the next scan window. If local/remote ratio is below
1281 * NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) the
1282 * scan period will decrease
1284 #define NUMA_PERIOD_SLOTS 10
1285 #define NUMA_PERIOD_THRESHOLD 3
1288 * Increase the scan period (slow down scanning) if the majority of
1289 * our memory is already on our local node, or if the majority of
1290 * the page accesses are shared with other processes.
1291 * Otherwise, decrease the scan period.
1293 static void update_task_scan_period(struct task_struct *p,
1294 unsigned long shared, unsigned long private)
1296 unsigned int period_slot;
1300 unsigned long remote = p->numa_faults_locality[0];
1301 unsigned long local = p->numa_faults_locality[1];
1304 * If there were no record hinting faults then either the task is
1305 * completely idle or all activity is areas that are not of interest
1306 * to automatic numa balancing. Scan slower
1308 if (local + shared == 0) {
1309 p->numa_scan_period = min(p->numa_scan_period_max,
1310 p->numa_scan_period << 1);
1312 p->mm->numa_next_scan = jiffies +
1313 msecs_to_jiffies(p->numa_scan_period);
1319 * Prepare to scale scan period relative to the current period.
1320 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1321 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1322 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1324 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1325 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1326 if (ratio >= NUMA_PERIOD_THRESHOLD) {
1327 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1330 diff = slot * period_slot;
1332 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1335 * Scale scan rate increases based on sharing. There is an
1336 * inverse relationship between the degree of sharing and
1337 * the adjustment made to the scanning period. Broadly
1338 * speaking the intent is that there is little point
1339 * scanning faster if shared accesses dominate as it may
1340 * simply bounce migrations uselessly
1342 period_slot = DIV_ROUND_UP(diff, NUMA_PERIOD_SLOTS);
1343 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared));
1344 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1347 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1348 task_scan_min(p), task_scan_max(p));
1349 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1352 static void task_numa_placement(struct task_struct *p)
1354 int seq, nid, max_nid = -1, max_group_nid = -1;
1355 unsigned long max_faults = 0, max_group_faults = 0;
1356 unsigned long fault_types[2] = { 0, 0 };
1357 spinlock_t *group_lock = NULL;
1359 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1360 if (p->numa_scan_seq == seq)
1362 p->numa_scan_seq = seq;
1363 p->numa_scan_period_max = task_scan_max(p);
1365 /* If the task is part of a group prevent parallel updates to group stats */
1366 if (p->numa_group) {
1367 group_lock = &p->numa_group->lock;
1368 spin_lock(group_lock);
1371 /* Find the node with the highest number of faults */
1372 for_each_online_node(nid) {
1373 unsigned long faults = 0, group_faults = 0;
1376 for (priv = 0; priv < 2; priv++) {
1379 i = task_faults_idx(nid, priv);
1380 diff = -p->numa_faults[i];
1382 /* Decay existing window, copy faults since last scan */
1383 p->numa_faults[i] >>= 1;
1384 p->numa_faults[i] += p->numa_faults_buffer[i];
1385 fault_types[priv] += p->numa_faults_buffer[i];
1386 p->numa_faults_buffer[i] = 0;
1388 faults += p->numa_faults[i];
1389 diff += p->numa_faults[i];
1390 p->total_numa_faults += diff;
1391 if (p->numa_group) {
1392 /* safe because we can only change our own group */
1393 p->numa_group->faults[i] += diff;
1394 p->numa_group->total_faults += diff;
1395 group_faults += p->numa_group->faults[i];
1399 if (faults > max_faults) {
1400 max_faults = faults;
1404 if (group_faults > max_group_faults) {
1405 max_group_faults = group_faults;
1406 max_group_nid = nid;
1410 update_task_scan_period(p, fault_types[0], fault_types[1]);
1412 if (p->numa_group) {
1414 * If the preferred task and group nids are different,
1415 * iterate over the nodes again to find the best place.
1417 if (max_nid != max_group_nid) {
1418 unsigned long weight, max_weight = 0;
1420 for_each_online_node(nid) {
1421 weight = task_weight(p, nid) + group_weight(p, nid);
1422 if (weight > max_weight) {
1423 max_weight = weight;
1429 spin_unlock(group_lock);
1432 /* Preferred node as the node with the most faults */
1433 if (max_faults && max_nid != p->numa_preferred_nid) {
1434 /* Update the preferred nid and migrate task if possible */
1435 sched_setnuma(p, max_nid);
1436 numa_migrate_preferred(p);
1440 static inline int get_numa_group(struct numa_group *grp)
1442 return atomic_inc_not_zero(&grp->refcount);
1445 static inline void put_numa_group(struct numa_group *grp)
1447 if (atomic_dec_and_test(&grp->refcount))
1448 kfree_rcu(grp, rcu);
1451 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1457 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1460 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
1463 struct numa_group *grp, *my_grp;
1464 struct task_struct *tsk;
1466 int cpu = cpupid_to_cpu(cpupid);
1469 if (unlikely(!p->numa_group)) {
1470 unsigned int size = sizeof(struct numa_group) +
1471 2*nr_node_ids*sizeof(unsigned long);
1473 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1477 atomic_set(&grp->refcount, 1);
1478 spin_lock_init(&grp->lock);
1479 INIT_LIST_HEAD(&grp->task_list);
1482 for (i = 0; i < 2*nr_node_ids; i++)
1483 grp->faults[i] = p->numa_faults[i];
1485 grp->total_faults = p->total_numa_faults;
1487 list_add(&p->numa_entry, &grp->task_list);
1489 rcu_assign_pointer(p->numa_group, grp);
1493 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1495 if (!cpupid_match_pid(tsk, cpupid))
1498 grp = rcu_dereference(tsk->numa_group);
1502 my_grp = p->numa_group;
1507 * Only join the other group if its bigger; if we're the bigger group,
1508 * the other task will join us.
1510 if (my_grp->nr_tasks > grp->nr_tasks)
1514 * Tie-break on the grp address.
1516 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1519 /* Always join threads in the same process. */
1520 if (tsk->mm == current->mm)
1523 /* Simple filter to avoid false positives due to PID collisions */
1524 if (flags & TNF_SHARED)
1527 /* Update priv based on whether false sharing was detected */
1530 if (join && !get_numa_group(grp))
1539 double_lock(&my_grp->lock, &grp->lock);
1541 for (i = 0; i < 2*nr_node_ids; i++) {
1542 my_grp->faults[i] -= p->numa_faults[i];
1543 grp->faults[i] += p->numa_faults[i];
1545 my_grp->total_faults -= p->total_numa_faults;
1546 grp->total_faults += p->total_numa_faults;
1548 list_move(&p->numa_entry, &grp->task_list);
1552 spin_unlock(&my_grp->lock);
1553 spin_unlock(&grp->lock);
1555 rcu_assign_pointer(p->numa_group, grp);
1557 put_numa_group(my_grp);
1560 void task_numa_free(struct task_struct *p)
1562 struct numa_group *grp = p->numa_group;
1564 void *numa_faults = p->numa_faults;
1567 spin_lock(&grp->lock);
1568 for (i = 0; i < 2*nr_node_ids; i++)
1569 grp->faults[i] -= p->numa_faults[i];
1570 grp->total_faults -= p->total_numa_faults;
1572 list_del(&p->numa_entry);
1574 spin_unlock(&grp->lock);
1575 rcu_assign_pointer(p->numa_group, NULL);
1576 put_numa_group(grp);
1579 p->numa_faults = NULL;
1580 p->numa_faults_buffer = NULL;
1585 * Got a PROT_NONE fault for a page on @node.
1587 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1589 struct task_struct *p = current;
1590 bool migrated = flags & TNF_MIGRATED;
1593 if (!numabalancing_enabled)
1596 /* for example, ksmd faulting in a user's mm */
1600 /* Do not worry about placement if exiting */
1601 if (p->state == TASK_DEAD)
1604 /* Allocate buffer to track faults on a per-node basis */
1605 if (unlikely(!p->numa_faults)) {
1606 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1608 /* numa_faults and numa_faults_buffer share the allocation */
1609 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1610 if (!p->numa_faults)
1613 BUG_ON(p->numa_faults_buffer);
1614 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1615 p->total_numa_faults = 0;
1616 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1620 * First accesses are treated as private, otherwise consider accesses
1621 * to be private if the accessing pid has not changed
1623 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1626 priv = cpupid_match_pid(p, last_cpupid);
1627 if (!priv && !(flags & TNF_NO_GROUP))
1628 task_numa_group(p, last_cpupid, flags, &priv);
1631 task_numa_placement(p);
1634 * Retry task to preferred node migration periodically, in case it
1635 * case it previously failed, or the scheduler moved us.
1637 if (time_after(jiffies, p->numa_migrate_retry))
1638 numa_migrate_preferred(p);
1641 p->numa_pages_migrated += pages;
1643 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1644 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1647 static void reset_ptenuma_scan(struct task_struct *p)
1649 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1650 p->mm->numa_scan_offset = 0;
1654 * The expensive part of numa migration is done from task_work context.
1655 * Triggered from task_tick_numa().
1657 void task_numa_work(struct callback_head *work)
1659 unsigned long migrate, next_scan, now = jiffies;
1660 struct task_struct *p = current;
1661 struct mm_struct *mm = p->mm;
1662 struct vm_area_struct *vma;
1663 unsigned long start, end;
1664 unsigned long nr_pte_updates = 0;
1667 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1669 work->next = work; /* protect against double add */
1671 * Who cares about NUMA placement when they're dying.
1673 * NOTE: make sure not to dereference p->mm before this check,
1674 * exit_task_work() happens _after_ exit_mm() so we could be called
1675 * without p->mm even though we still had it when we enqueued this
1678 if (p->flags & PF_EXITING)
1681 if (!mm->numa_next_scan) {
1682 mm->numa_next_scan = now +
1683 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1687 * Enforce maximal scan/migration frequency..
1689 migrate = mm->numa_next_scan;
1690 if (time_before(now, migrate))
1693 if (p->numa_scan_period == 0) {
1694 p->numa_scan_period_max = task_scan_max(p);
1695 p->numa_scan_period = task_scan_min(p);
1698 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1699 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1703 * Delay this task enough that another task of this mm will likely win
1704 * the next time around.
1706 p->node_stamp += 2 * TICK_NSEC;
1708 start = mm->numa_scan_offset;
1709 pages = sysctl_numa_balancing_scan_size;
1710 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1714 down_read(&mm->mmap_sem);
1715 vma = find_vma(mm, start);
1717 reset_ptenuma_scan(p);
1721 for (; vma; vma = vma->vm_next) {
1722 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1726 * Shared library pages mapped by multiple processes are not
1727 * migrated as it is expected they are cache replicated. Avoid
1728 * hinting faults in read-only file-backed mappings or the vdso
1729 * as migrating the pages will be of marginal benefit.
1732 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1736 start = max(start, vma->vm_start);
1737 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1738 end = min(end, vma->vm_end);
1739 nr_pte_updates += change_prot_numa(vma, start, end);
1742 * Scan sysctl_numa_balancing_scan_size but ensure that
1743 * at least one PTE is updated so that unused virtual
1744 * address space is quickly skipped.
1747 pages -= (end - start) >> PAGE_SHIFT;
1752 } while (end != vma->vm_end);
1757 * It is possible to reach the end of the VMA list but the last few
1758 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1759 * would find the !migratable VMA on the next scan but not reset the
1760 * scanner to the start so check it now.
1763 mm->numa_scan_offset = start;
1765 reset_ptenuma_scan(p);
1766 up_read(&mm->mmap_sem);
1770 * Drive the periodic memory faults..
1772 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1774 struct callback_head *work = &curr->numa_work;
1778 * We don't care about NUMA placement if we don't have memory.
1780 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1784 * Using runtime rather than walltime has the dual advantage that
1785 * we (mostly) drive the selection from busy threads and that the
1786 * task needs to have done some actual work before we bother with
1789 now = curr->se.sum_exec_runtime;
1790 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1792 if (now - curr->node_stamp > period) {
1793 if (!curr->node_stamp)
1794 curr->numa_scan_period = task_scan_min(curr);
1795 curr->node_stamp += period;
1797 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1798 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1799 task_work_add(curr, work, true);
1804 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1808 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1812 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1815 #endif /* CONFIG_NUMA_BALANCING */
1818 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1820 update_load_add(&cfs_rq->load, se->load.weight);
1821 if (!parent_entity(se))
1822 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1824 if (entity_is_task(se)) {
1825 struct rq *rq = rq_of(cfs_rq);
1827 account_numa_enqueue(rq, task_of(se));
1828 list_add(&se->group_node, &rq->cfs_tasks);
1831 cfs_rq->nr_running++;
1835 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1837 update_load_sub(&cfs_rq->load, se->load.weight);
1838 if (!parent_entity(se))
1839 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1840 if (entity_is_task(se)) {
1841 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1842 list_del_init(&se->group_node);
1844 cfs_rq->nr_running--;
1847 #ifdef CONFIG_FAIR_GROUP_SCHED
1849 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1854 * Use this CPU's actual weight instead of the last load_contribution
1855 * to gain a more accurate current total weight. See
1856 * update_cfs_rq_load_contribution().
1858 tg_weight = atomic_long_read(&tg->load_avg);
1859 tg_weight -= cfs_rq->tg_load_contrib;
1860 tg_weight += cfs_rq->load.weight;
1865 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1867 long tg_weight, load, shares;
1869 tg_weight = calc_tg_weight(tg, cfs_rq);
1870 load = cfs_rq->load.weight;
1872 shares = (tg->shares * load);
1874 shares /= tg_weight;
1876 if (shares < MIN_SHARES)
1877 shares = MIN_SHARES;
1878 if (shares > tg->shares)
1879 shares = tg->shares;
1883 # else /* CONFIG_SMP */
1884 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1888 # endif /* CONFIG_SMP */
1889 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1890 unsigned long weight)
1893 /* commit outstanding execution time */
1894 if (cfs_rq->curr == se)
1895 update_curr(cfs_rq);
1896 account_entity_dequeue(cfs_rq, se);
1899 update_load_set(&se->load, weight);
1902 account_entity_enqueue(cfs_rq, se);
1905 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1907 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1909 struct task_group *tg;
1910 struct sched_entity *se;
1914 se = tg->se[cpu_of(rq_of(cfs_rq))];
1915 if (!se || throttled_hierarchy(cfs_rq))
1918 if (likely(se->load.weight == tg->shares))
1921 shares = calc_cfs_shares(cfs_rq, tg);
1923 reweight_entity(cfs_rq_of(se), se, shares);
1925 #else /* CONFIG_FAIR_GROUP_SCHED */
1926 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1929 #endif /* CONFIG_FAIR_GROUP_SCHED */
1933 * We choose a half-life close to 1 scheduling period.
1934 * Note: The tables below are dependent on this value.
1936 #define LOAD_AVG_PERIOD 32
1937 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1938 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1940 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1941 static const u32 runnable_avg_yN_inv[] = {
1942 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1943 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1944 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1945 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1946 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1947 0x85aac367, 0x82cd8698,
1951 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1952 * over-estimates when re-combining.
1954 static const u32 runnable_avg_yN_sum[] = {
1955 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1956 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1957 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1962 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1964 static __always_inline u64 decay_load(u64 val, u64 n)
1966 unsigned int local_n;
1970 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1973 /* after bounds checking we can collapse to 32-bit */
1977 * As y^PERIOD = 1/2, we can combine
1978 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1979 * With a look-up table which covers k^n (n<PERIOD)
1981 * To achieve constant time decay_load.
1983 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1984 val >>= local_n / LOAD_AVG_PERIOD;
1985 local_n %= LOAD_AVG_PERIOD;
1988 val *= runnable_avg_yN_inv[local_n];
1989 /* We don't use SRR here since we always want to round down. */
1994 * For updates fully spanning n periods, the contribution to runnable
1995 * average will be: \Sum 1024*y^n
1997 * We can compute this reasonably efficiently by combining:
1998 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2000 static u32 __compute_runnable_contrib(u64 n)
2004 if (likely(n <= LOAD_AVG_PERIOD))
2005 return runnable_avg_yN_sum[n];
2006 else if (unlikely(n >= LOAD_AVG_MAX_N))
2007 return LOAD_AVG_MAX;
2009 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2011 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2012 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2014 n -= LOAD_AVG_PERIOD;
2015 } while (n > LOAD_AVG_PERIOD);
2017 contrib = decay_load(contrib, n);
2018 return contrib + runnable_avg_yN_sum[n];
2022 * We can represent the historical contribution to runnable average as the
2023 * coefficients of a geometric series. To do this we sub-divide our runnable
2024 * history into segments of approximately 1ms (1024us); label the segment that
2025 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2027 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2029 * (now) (~1ms ago) (~2ms ago)
2031 * Let u_i denote the fraction of p_i that the entity was runnable.
2033 * We then designate the fractions u_i as our co-efficients, yielding the
2034 * following representation of historical load:
2035 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2037 * We choose y based on the with of a reasonably scheduling period, fixing:
2040 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2041 * approximately half as much as the contribution to load within the last ms
2044 * When a period "rolls over" and we have new u_0`, multiplying the previous
2045 * sum again by y is sufficient to update:
2046 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2047 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2049 static __always_inline int __update_entity_runnable_avg(u64 now,
2050 struct sched_avg *sa,
2054 u32 runnable_contrib;
2055 int delta_w, decayed = 0;
2057 delta = now - sa->last_runnable_update;
2059 * This should only happen when time goes backwards, which it
2060 * unfortunately does during sched clock init when we swap over to TSC.
2062 if ((s64)delta < 0) {
2063 sa->last_runnable_update = now;
2068 * Use 1024ns as the unit of measurement since it's a reasonable
2069 * approximation of 1us and fast to compute.
2074 sa->last_runnable_update = now;
2076 /* delta_w is the amount already accumulated against our next period */
2077 delta_w = sa->runnable_avg_period % 1024;
2078 if (delta + delta_w >= 1024) {
2079 /* period roll-over */
2083 * Now that we know we're crossing a period boundary, figure
2084 * out how much from delta we need to complete the current
2085 * period and accrue it.
2087 delta_w = 1024 - delta_w;
2089 sa->runnable_avg_sum += delta_w;
2090 sa->runnable_avg_period += delta_w;
2094 /* Figure out how many additional periods this update spans */
2095 periods = delta / 1024;
2098 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2100 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2103 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2104 runnable_contrib = __compute_runnable_contrib(periods);
2106 sa->runnable_avg_sum += runnable_contrib;
2107 sa->runnable_avg_period += runnable_contrib;
2110 /* Remainder of delta accrued against u_0` */
2112 sa->runnable_avg_sum += delta;
2113 sa->runnable_avg_period += delta;
2118 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2119 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2121 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2122 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2124 decays -= se->avg.decay_count;
2128 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2129 se->avg.decay_count = 0;
2134 #ifdef CONFIG_FAIR_GROUP_SCHED
2135 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2138 struct task_group *tg = cfs_rq->tg;
2141 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2142 tg_contrib -= cfs_rq->tg_load_contrib;
2144 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2145 atomic_long_add(tg_contrib, &tg->load_avg);
2146 cfs_rq->tg_load_contrib += tg_contrib;
2151 * Aggregate cfs_rq runnable averages into an equivalent task_group
2152 * representation for computing load contributions.
2154 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2155 struct cfs_rq *cfs_rq)
2157 struct task_group *tg = cfs_rq->tg;
2160 /* The fraction of a cpu used by this cfs_rq */
2161 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2162 sa->runnable_avg_period + 1);
2163 contrib -= cfs_rq->tg_runnable_contrib;
2165 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2166 atomic_add(contrib, &tg->runnable_avg);
2167 cfs_rq->tg_runnable_contrib += contrib;
2171 static inline void __update_group_entity_contrib(struct sched_entity *se)
2173 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2174 struct task_group *tg = cfs_rq->tg;
2179 contrib = cfs_rq->tg_load_contrib * tg->shares;
2180 se->avg.load_avg_contrib = div_u64(contrib,
2181 atomic_long_read(&tg->load_avg) + 1);
2184 * For group entities we need to compute a correction term in the case
2185 * that they are consuming <1 cpu so that we would contribute the same
2186 * load as a task of equal weight.
2188 * Explicitly co-ordinating this measurement would be expensive, but
2189 * fortunately the sum of each cpus contribution forms a usable
2190 * lower-bound on the true value.
2192 * Consider the aggregate of 2 contributions. Either they are disjoint
2193 * (and the sum represents true value) or they are disjoint and we are
2194 * understating by the aggregate of their overlap.
2196 * Extending this to N cpus, for a given overlap, the maximum amount we
2197 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2198 * cpus that overlap for this interval and w_i is the interval width.
2200 * On a small machine; the first term is well-bounded which bounds the
2201 * total error since w_i is a subset of the period. Whereas on a
2202 * larger machine, while this first term can be larger, if w_i is the
2203 * of consequential size guaranteed to see n_i*w_i quickly converge to
2204 * our upper bound of 1-cpu.
2206 runnable_avg = atomic_read(&tg->runnable_avg);
2207 if (runnable_avg < NICE_0_LOAD) {
2208 se->avg.load_avg_contrib *= runnable_avg;
2209 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2213 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2214 int force_update) {}
2215 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2216 struct cfs_rq *cfs_rq) {}
2217 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2220 static inline void __update_task_entity_contrib(struct sched_entity *se)
2224 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2225 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2226 contrib /= (se->avg.runnable_avg_period + 1);
2227 se->avg.load_avg_contrib = scale_load(contrib);
2230 /* Compute the current contribution to load_avg by se, return any delta */
2231 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2233 long old_contrib = se->avg.load_avg_contrib;
2235 if (entity_is_task(se)) {
2236 __update_task_entity_contrib(se);
2238 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2239 __update_group_entity_contrib(se);
2242 return se->avg.load_avg_contrib - old_contrib;
2245 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2248 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2249 cfs_rq->blocked_load_avg -= load_contrib;
2251 cfs_rq->blocked_load_avg = 0;
2254 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2256 /* Update a sched_entity's runnable average */
2257 static inline void update_entity_load_avg(struct sched_entity *se,
2260 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2265 * For a group entity we need to use their owned cfs_rq_clock_task() in
2266 * case they are the parent of a throttled hierarchy.
2268 if (entity_is_task(se))
2269 now = cfs_rq_clock_task(cfs_rq);
2271 now = cfs_rq_clock_task(group_cfs_rq(se));
2273 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2276 contrib_delta = __update_entity_load_avg_contrib(se);
2282 cfs_rq->runnable_load_avg += contrib_delta;
2284 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2288 * Decay the load contributed by all blocked children and account this so that
2289 * their contribution may appropriately discounted when they wake up.
2291 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2293 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2296 decays = now - cfs_rq->last_decay;
2297 if (!decays && !force_update)
2300 if (atomic_long_read(&cfs_rq->removed_load)) {
2301 unsigned long removed_load;
2302 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2303 subtract_blocked_load_contrib(cfs_rq, removed_load);
2307 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2309 atomic64_add(decays, &cfs_rq->decay_counter);
2310 cfs_rq->last_decay = now;
2313 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2316 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2318 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2319 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2322 /* Add the load generated by se into cfs_rq's child load-average */
2323 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2324 struct sched_entity *se,
2328 * We track migrations using entity decay_count <= 0, on a wake-up
2329 * migration we use a negative decay count to track the remote decays
2330 * accumulated while sleeping.
2332 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2333 * are seen by enqueue_entity_load_avg() as a migration with an already
2334 * constructed load_avg_contrib.
2336 if (unlikely(se->avg.decay_count <= 0)) {
2337 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2338 if (se->avg.decay_count) {
2340 * In a wake-up migration we have to approximate the
2341 * time sleeping. This is because we can't synchronize
2342 * clock_task between the two cpus, and it is not
2343 * guaranteed to be read-safe. Instead, we can
2344 * approximate this using our carried decays, which are
2345 * explicitly atomically readable.
2347 se->avg.last_runnable_update -= (-se->avg.decay_count)
2349 update_entity_load_avg(se, 0);
2350 /* Indicate that we're now synchronized and on-rq */
2351 se->avg.decay_count = 0;
2356 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2357 * would have made count negative); we must be careful to avoid
2358 * double-accounting blocked time after synchronizing decays.
2360 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2364 /* migrated tasks did not contribute to our blocked load */
2366 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2367 update_entity_load_avg(se, 0);
2370 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2371 /* we force update consideration on load-balancer moves */
2372 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2376 * Remove se's load from this cfs_rq child load-average, if the entity is
2377 * transitioning to a blocked state we track its projected decay using
2380 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2381 struct sched_entity *se,
2384 update_entity_load_avg(se, 1);
2385 /* we force update consideration on load-balancer moves */
2386 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2388 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2390 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2391 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2392 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2396 * Update the rq's load with the elapsed running time before entering
2397 * idle. if the last scheduled task is not a CFS task, idle_enter will
2398 * be the only way to update the runnable statistic.
2400 void idle_enter_fair(struct rq *this_rq)
2402 update_rq_runnable_avg(this_rq, 1);
2406 * Update the rq's load with the elapsed idle time before a task is
2407 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2408 * be the only way to update the runnable statistic.
2410 void idle_exit_fair(struct rq *this_rq)
2412 update_rq_runnable_avg(this_rq, 0);
2416 static inline void update_entity_load_avg(struct sched_entity *se,
2417 int update_cfs_rq) {}
2418 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2419 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2420 struct sched_entity *se,
2422 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2423 struct sched_entity *se,
2425 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2426 int force_update) {}
2429 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2431 #ifdef CONFIG_SCHEDSTATS
2432 struct task_struct *tsk = NULL;
2434 if (entity_is_task(se))
2437 if (se->statistics.sleep_start) {
2438 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2443 if (unlikely(delta > se->statistics.sleep_max))
2444 se->statistics.sleep_max = delta;
2446 se->statistics.sleep_start = 0;
2447 se->statistics.sum_sleep_runtime += delta;
2450 account_scheduler_latency(tsk, delta >> 10, 1);
2451 trace_sched_stat_sleep(tsk, delta);
2454 if (se->statistics.block_start) {
2455 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2460 if (unlikely(delta > se->statistics.block_max))
2461 se->statistics.block_max = delta;
2463 se->statistics.block_start = 0;
2464 se->statistics.sum_sleep_runtime += delta;
2467 if (tsk->in_iowait) {
2468 se->statistics.iowait_sum += delta;
2469 se->statistics.iowait_count++;
2470 trace_sched_stat_iowait(tsk, delta);
2473 trace_sched_stat_blocked(tsk, delta);
2476 * Blocking time is in units of nanosecs, so shift by
2477 * 20 to get a milliseconds-range estimation of the
2478 * amount of time that the task spent sleeping:
2480 if (unlikely(prof_on == SLEEP_PROFILING)) {
2481 profile_hits(SLEEP_PROFILING,
2482 (void *)get_wchan(tsk),
2485 account_scheduler_latency(tsk, delta >> 10, 0);
2491 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2493 #ifdef CONFIG_SCHED_DEBUG
2494 s64 d = se->vruntime - cfs_rq->min_vruntime;
2499 if (d > 3*sysctl_sched_latency)
2500 schedstat_inc(cfs_rq, nr_spread_over);
2505 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2507 u64 vruntime = cfs_rq->min_vruntime;
2510 * The 'current' period is already promised to the current tasks,
2511 * however the extra weight of the new task will slow them down a
2512 * little, place the new task so that it fits in the slot that
2513 * stays open at the end.
2515 if (initial && sched_feat(START_DEBIT))
2516 vruntime += sched_vslice(cfs_rq, se);
2518 /* sleeps up to a single latency don't count. */
2520 unsigned long thresh = sysctl_sched_latency;
2523 * Halve their sleep time's effect, to allow
2524 * for a gentler effect of sleepers:
2526 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2532 /* ensure we never gain time by being placed backwards. */
2533 se->vruntime = max_vruntime(se->vruntime, vruntime);
2536 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2539 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2542 * Update the normalized vruntime before updating min_vruntime
2543 * through calling update_curr().
2545 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2546 se->vruntime += cfs_rq->min_vruntime;
2549 * Update run-time statistics of the 'current'.
2551 update_curr(cfs_rq);
2552 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2553 account_entity_enqueue(cfs_rq, se);
2554 update_cfs_shares(cfs_rq);
2556 if (flags & ENQUEUE_WAKEUP) {
2557 place_entity(cfs_rq, se, 0);
2558 enqueue_sleeper(cfs_rq, se);
2561 update_stats_enqueue(cfs_rq, se);
2562 check_spread(cfs_rq, se);
2563 if (se != cfs_rq->curr)
2564 __enqueue_entity(cfs_rq, se);
2567 if (cfs_rq->nr_running == 1) {
2568 list_add_leaf_cfs_rq(cfs_rq);
2569 check_enqueue_throttle(cfs_rq);
2573 static void __clear_buddies_last(struct sched_entity *se)
2575 for_each_sched_entity(se) {
2576 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2577 if (cfs_rq->last == se)
2578 cfs_rq->last = NULL;
2584 static void __clear_buddies_next(struct sched_entity *se)
2586 for_each_sched_entity(se) {
2587 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2588 if (cfs_rq->next == se)
2589 cfs_rq->next = NULL;
2595 static void __clear_buddies_skip(struct sched_entity *se)
2597 for_each_sched_entity(se) {
2598 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2599 if (cfs_rq->skip == se)
2600 cfs_rq->skip = NULL;
2606 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2608 if (cfs_rq->last == se)
2609 __clear_buddies_last(se);
2611 if (cfs_rq->next == se)
2612 __clear_buddies_next(se);
2614 if (cfs_rq->skip == se)
2615 __clear_buddies_skip(se);
2618 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2621 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2624 * Update run-time statistics of the 'current'.
2626 update_curr(cfs_rq);
2627 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2629 update_stats_dequeue(cfs_rq, se);
2630 if (flags & DEQUEUE_SLEEP) {
2631 #ifdef CONFIG_SCHEDSTATS
2632 if (entity_is_task(se)) {
2633 struct task_struct *tsk = task_of(se);
2635 if (tsk->state & TASK_INTERRUPTIBLE)
2636 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2637 if (tsk->state & TASK_UNINTERRUPTIBLE)
2638 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2643 clear_buddies(cfs_rq, se);
2645 if (se != cfs_rq->curr)
2646 __dequeue_entity(cfs_rq, se);
2648 account_entity_dequeue(cfs_rq, se);
2651 * Normalize the entity after updating the min_vruntime because the
2652 * update can refer to the ->curr item and we need to reflect this
2653 * movement in our normalized position.
2655 if (!(flags & DEQUEUE_SLEEP))
2656 se->vruntime -= cfs_rq->min_vruntime;
2658 /* return excess runtime on last dequeue */
2659 return_cfs_rq_runtime(cfs_rq);
2661 update_min_vruntime(cfs_rq);
2662 update_cfs_shares(cfs_rq);
2666 * Preempt the current task with a newly woken task if needed:
2669 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2671 unsigned long ideal_runtime, delta_exec;
2672 struct sched_entity *se;
2675 ideal_runtime = sched_slice(cfs_rq, curr);
2676 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2677 if (delta_exec > ideal_runtime) {
2678 resched_task(rq_of(cfs_rq)->curr);
2680 * The current task ran long enough, ensure it doesn't get
2681 * re-elected due to buddy favours.
2683 clear_buddies(cfs_rq, curr);
2688 * Ensure that a task that missed wakeup preemption by a
2689 * narrow margin doesn't have to wait for a full slice.
2690 * This also mitigates buddy induced latencies under load.
2692 if (delta_exec < sysctl_sched_min_granularity)
2695 se = __pick_first_entity(cfs_rq);
2696 delta = curr->vruntime - se->vruntime;
2701 if (delta > ideal_runtime)
2702 resched_task(rq_of(cfs_rq)->curr);
2706 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2708 /* 'current' is not kept within the tree. */
2711 * Any task has to be enqueued before it get to execute on
2712 * a CPU. So account for the time it spent waiting on the
2715 update_stats_wait_end(cfs_rq, se);
2716 __dequeue_entity(cfs_rq, se);
2719 update_stats_curr_start(cfs_rq, se);
2721 #ifdef CONFIG_SCHEDSTATS
2723 * Track our maximum slice length, if the CPU's load is at
2724 * least twice that of our own weight (i.e. dont track it
2725 * when there are only lesser-weight tasks around):
2727 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2728 se->statistics.slice_max = max(se->statistics.slice_max,
2729 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2732 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2736 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2739 * Pick the next process, keeping these things in mind, in this order:
2740 * 1) keep things fair between processes/task groups
2741 * 2) pick the "next" process, since someone really wants that to run
2742 * 3) pick the "last" process, for cache locality
2743 * 4) do not run the "skip" process, if something else is available
2745 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2747 struct sched_entity *se = __pick_first_entity(cfs_rq);
2748 struct sched_entity *left = se;
2751 * Avoid running the skip buddy, if running something else can
2752 * be done without getting too unfair.
2754 if (cfs_rq->skip == se) {
2755 struct sched_entity *second = __pick_next_entity(se);
2756 if (second && wakeup_preempt_entity(second, left) < 1)
2761 * Prefer last buddy, try to return the CPU to a preempted task.
2763 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2767 * Someone really wants this to run. If it's not unfair, run it.
2769 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2772 clear_buddies(cfs_rq, se);
2777 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2779 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2782 * If still on the runqueue then deactivate_task()
2783 * was not called and update_curr() has to be done:
2786 update_curr(cfs_rq);
2788 /* throttle cfs_rqs exceeding runtime */
2789 check_cfs_rq_runtime(cfs_rq);
2791 check_spread(cfs_rq, prev);
2793 update_stats_wait_start(cfs_rq, prev);
2794 /* Put 'current' back into the tree. */
2795 __enqueue_entity(cfs_rq, prev);
2796 /* in !on_rq case, update occurred at dequeue */
2797 update_entity_load_avg(prev, 1);
2799 cfs_rq->curr = NULL;
2803 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2806 * Update run-time statistics of the 'current'.
2808 update_curr(cfs_rq);
2811 * Ensure that runnable average is periodically updated.
2813 update_entity_load_avg(curr, 1);
2814 update_cfs_rq_blocked_load(cfs_rq, 1);
2815 update_cfs_shares(cfs_rq);
2817 #ifdef CONFIG_SCHED_HRTICK
2819 * queued ticks are scheduled to match the slice, so don't bother
2820 * validating it and just reschedule.
2823 resched_task(rq_of(cfs_rq)->curr);
2827 * don't let the period tick interfere with the hrtick preemption
2829 if (!sched_feat(DOUBLE_TICK) &&
2830 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2834 if (cfs_rq->nr_running > 1)
2835 check_preempt_tick(cfs_rq, curr);
2839 /**************************************************
2840 * CFS bandwidth control machinery
2843 #ifdef CONFIG_CFS_BANDWIDTH
2845 #ifdef HAVE_JUMP_LABEL
2846 static struct static_key __cfs_bandwidth_used;
2848 static inline bool cfs_bandwidth_used(void)
2850 return static_key_false(&__cfs_bandwidth_used);
2853 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2855 /* only need to count groups transitioning between enabled/!enabled */
2856 if (enabled && !was_enabled)
2857 static_key_slow_inc(&__cfs_bandwidth_used);
2858 else if (!enabled && was_enabled)
2859 static_key_slow_dec(&__cfs_bandwidth_used);
2861 #else /* HAVE_JUMP_LABEL */
2862 static bool cfs_bandwidth_used(void)
2867 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2868 #endif /* HAVE_JUMP_LABEL */
2871 * default period for cfs group bandwidth.
2872 * default: 0.1s, units: nanoseconds
2874 static inline u64 default_cfs_period(void)
2876 return 100000000ULL;
2879 static inline u64 sched_cfs_bandwidth_slice(void)
2881 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2885 * Replenish runtime according to assigned quota and update expiration time.
2886 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2887 * additional synchronization around rq->lock.
2889 * requires cfs_b->lock
2891 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2895 if (cfs_b->quota == RUNTIME_INF)
2898 now = sched_clock_cpu(smp_processor_id());
2899 cfs_b->runtime = cfs_b->quota;
2900 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2903 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2905 return &tg->cfs_bandwidth;
2908 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2909 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2911 if (unlikely(cfs_rq->throttle_count))
2912 return cfs_rq->throttled_clock_task;
2914 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2917 /* returns 0 on failure to allocate runtime */
2918 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2920 struct task_group *tg = cfs_rq->tg;
2921 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2922 u64 amount = 0, min_amount, expires;
2924 /* note: this is a positive sum as runtime_remaining <= 0 */
2925 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2927 raw_spin_lock(&cfs_b->lock);
2928 if (cfs_b->quota == RUNTIME_INF)
2929 amount = min_amount;
2932 * If the bandwidth pool has become inactive, then at least one
2933 * period must have elapsed since the last consumption.
2934 * Refresh the global state and ensure bandwidth timer becomes
2937 if (!cfs_b->timer_active) {
2938 __refill_cfs_bandwidth_runtime(cfs_b);
2939 __start_cfs_bandwidth(cfs_b);
2942 if (cfs_b->runtime > 0) {
2943 amount = min(cfs_b->runtime, min_amount);
2944 cfs_b->runtime -= amount;
2948 expires = cfs_b->runtime_expires;
2949 raw_spin_unlock(&cfs_b->lock);
2951 cfs_rq->runtime_remaining += amount;
2953 * we may have advanced our local expiration to account for allowed
2954 * spread between our sched_clock and the one on which runtime was
2957 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2958 cfs_rq->runtime_expires = expires;
2960 return cfs_rq->runtime_remaining > 0;
2964 * Note: This depends on the synchronization provided by sched_clock and the
2965 * fact that rq->clock snapshots this value.
2967 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2969 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2971 /* if the deadline is ahead of our clock, nothing to do */
2972 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2975 if (cfs_rq->runtime_remaining < 0)
2979 * If the local deadline has passed we have to consider the
2980 * possibility that our sched_clock is 'fast' and the global deadline
2981 * has not truly expired.
2983 * Fortunately we can check determine whether this the case by checking
2984 * whether the global deadline has advanced.
2987 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2988 /* extend local deadline, drift is bounded above by 2 ticks */
2989 cfs_rq->runtime_expires += TICK_NSEC;
2991 /* global deadline is ahead, expiration has passed */
2992 cfs_rq->runtime_remaining = 0;
2996 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2997 unsigned long delta_exec)
2999 /* dock delta_exec before expiring quota (as it could span periods) */
3000 cfs_rq->runtime_remaining -= delta_exec;
3001 expire_cfs_rq_runtime(cfs_rq);
3003 if (likely(cfs_rq->runtime_remaining > 0))
3007 * if we're unable to extend our runtime we resched so that the active
3008 * hierarchy can be throttled
3010 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3011 resched_task(rq_of(cfs_rq)->curr);
3014 static __always_inline
3015 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
3017 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3020 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3023 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3025 return cfs_bandwidth_used() && cfs_rq->throttled;
3028 /* check whether cfs_rq, or any parent, is throttled */
3029 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3031 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3035 * Ensure that neither of the group entities corresponding to src_cpu or
3036 * dest_cpu are members of a throttled hierarchy when performing group
3037 * load-balance operations.
3039 static inline int throttled_lb_pair(struct task_group *tg,
3040 int src_cpu, int dest_cpu)
3042 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3044 src_cfs_rq = tg->cfs_rq[src_cpu];
3045 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3047 return throttled_hierarchy(src_cfs_rq) ||
3048 throttled_hierarchy(dest_cfs_rq);
3051 /* updated child weight may affect parent so we have to do this bottom up */
3052 static int tg_unthrottle_up(struct task_group *tg, void *data)
3054 struct rq *rq = data;
3055 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3057 cfs_rq->throttle_count--;
3059 if (!cfs_rq->throttle_count) {
3060 /* adjust cfs_rq_clock_task() */
3061 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3062 cfs_rq->throttled_clock_task;
3069 static int tg_throttle_down(struct task_group *tg, void *data)
3071 struct rq *rq = data;
3072 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3074 /* group is entering throttled state, stop time */
3075 if (!cfs_rq->throttle_count)
3076 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3077 cfs_rq->throttle_count++;
3082 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3084 struct rq *rq = rq_of(cfs_rq);
3085 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3086 struct sched_entity *se;
3087 long task_delta, dequeue = 1;
3089 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3091 /* freeze hierarchy runnable averages while throttled */
3093 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3096 task_delta = cfs_rq->h_nr_running;
3097 for_each_sched_entity(se) {
3098 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3099 /* throttled entity or throttle-on-deactivate */
3104 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3105 qcfs_rq->h_nr_running -= task_delta;
3107 if (qcfs_rq->load.weight)
3112 rq->nr_running -= task_delta;
3114 cfs_rq->throttled = 1;
3115 cfs_rq->throttled_clock = rq_clock(rq);
3116 raw_spin_lock(&cfs_b->lock);
3117 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3118 raw_spin_unlock(&cfs_b->lock);
3121 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3123 struct rq *rq = rq_of(cfs_rq);
3124 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3125 struct sched_entity *se;
3129 se = cfs_rq->tg->se[cpu_of(rq)];
3131 cfs_rq->throttled = 0;
3133 update_rq_clock(rq);
3135 raw_spin_lock(&cfs_b->lock);
3136 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3137 list_del_rcu(&cfs_rq->throttled_list);
3138 raw_spin_unlock(&cfs_b->lock);
3140 /* update hierarchical throttle state */
3141 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3143 if (!cfs_rq->load.weight)
3146 task_delta = cfs_rq->h_nr_running;
3147 for_each_sched_entity(se) {
3151 cfs_rq = cfs_rq_of(se);
3153 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3154 cfs_rq->h_nr_running += task_delta;
3156 if (cfs_rq_throttled(cfs_rq))
3161 rq->nr_running += task_delta;
3163 /* determine whether we need to wake up potentially idle cpu */
3164 if (rq->curr == rq->idle && rq->cfs.nr_running)
3165 resched_task(rq->curr);
3168 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3169 u64 remaining, u64 expires)
3171 struct cfs_rq *cfs_rq;
3172 u64 runtime = remaining;
3175 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3177 struct rq *rq = rq_of(cfs_rq);
3179 raw_spin_lock(&rq->lock);
3180 if (!cfs_rq_throttled(cfs_rq))
3183 runtime = -cfs_rq->runtime_remaining + 1;
3184 if (runtime > remaining)
3185 runtime = remaining;
3186 remaining -= runtime;
3188 cfs_rq->runtime_remaining += runtime;
3189 cfs_rq->runtime_expires = expires;
3191 /* we check whether we're throttled above */
3192 if (cfs_rq->runtime_remaining > 0)
3193 unthrottle_cfs_rq(cfs_rq);
3196 raw_spin_unlock(&rq->lock);
3207 * Responsible for refilling a task_group's bandwidth and unthrottling its
3208 * cfs_rqs as appropriate. If there has been no activity within the last
3209 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3210 * used to track this state.
3212 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3214 u64 runtime, runtime_expires;
3215 int idle = 1, throttled;
3217 raw_spin_lock(&cfs_b->lock);
3218 /* no need to continue the timer with no bandwidth constraint */
3219 if (cfs_b->quota == RUNTIME_INF)
3222 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3223 /* idle depends on !throttled (for the case of a large deficit) */
3224 idle = cfs_b->idle && !throttled;
3225 cfs_b->nr_periods += overrun;
3227 /* if we're going inactive then everything else can be deferred */
3231 __refill_cfs_bandwidth_runtime(cfs_b);
3234 /* mark as potentially idle for the upcoming period */
3239 /* account preceding periods in which throttling occurred */
3240 cfs_b->nr_throttled += overrun;
3243 * There are throttled entities so we must first use the new bandwidth
3244 * to unthrottle them before making it generally available. This
3245 * ensures that all existing debts will be paid before a new cfs_rq is
3248 runtime = cfs_b->runtime;
3249 runtime_expires = cfs_b->runtime_expires;
3253 * This check is repeated as we are holding onto the new bandwidth
3254 * while we unthrottle. This can potentially race with an unthrottled
3255 * group trying to acquire new bandwidth from the global pool.
3257 while (throttled && runtime > 0) {
3258 raw_spin_unlock(&cfs_b->lock);
3259 /* we can't nest cfs_b->lock while distributing bandwidth */
3260 runtime = distribute_cfs_runtime(cfs_b, runtime,
3262 raw_spin_lock(&cfs_b->lock);
3264 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3267 /* return (any) remaining runtime */
3268 cfs_b->runtime = runtime;
3270 * While we are ensured activity in the period following an
3271 * unthrottle, this also covers the case in which the new bandwidth is
3272 * insufficient to cover the existing bandwidth deficit. (Forcing the
3273 * timer to remain active while there are any throttled entities.)
3278 cfs_b->timer_active = 0;
3279 raw_spin_unlock(&cfs_b->lock);
3284 /* a cfs_rq won't donate quota below this amount */
3285 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3286 /* minimum remaining period time to redistribute slack quota */
3287 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3288 /* how long we wait to gather additional slack before distributing */
3289 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3291 /* are we near the end of the current quota period? */
3292 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3294 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3297 /* if the call-back is running a quota refresh is already occurring */
3298 if (hrtimer_callback_running(refresh_timer))
3301 /* is a quota refresh about to occur? */
3302 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3303 if (remaining < min_expire)
3309 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3311 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3313 /* if there's a quota refresh soon don't bother with slack */
3314 if (runtime_refresh_within(cfs_b, min_left))
3317 start_bandwidth_timer(&cfs_b->slack_timer,
3318 ns_to_ktime(cfs_bandwidth_slack_period));
3321 /* we know any runtime found here is valid as update_curr() precedes return */
3322 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3324 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3325 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3327 if (slack_runtime <= 0)
3330 raw_spin_lock(&cfs_b->lock);
3331 if (cfs_b->quota != RUNTIME_INF &&
3332 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3333 cfs_b->runtime += slack_runtime;
3335 /* we are under rq->lock, defer unthrottling using a timer */
3336 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3337 !list_empty(&cfs_b->throttled_cfs_rq))
3338 start_cfs_slack_bandwidth(cfs_b);
3340 raw_spin_unlock(&cfs_b->lock);
3342 /* even if it's not valid for return we don't want to try again */
3343 cfs_rq->runtime_remaining -= slack_runtime;
3346 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3348 if (!cfs_bandwidth_used())
3351 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3354 __return_cfs_rq_runtime(cfs_rq);
3358 * This is done with a timer (instead of inline with bandwidth return) since
3359 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3361 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3363 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3366 /* confirm we're still not at a refresh boundary */
3367 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3370 raw_spin_lock(&cfs_b->lock);
3371 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3372 runtime = cfs_b->runtime;
3375 expires = cfs_b->runtime_expires;
3376 raw_spin_unlock(&cfs_b->lock);
3381 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3383 raw_spin_lock(&cfs_b->lock);
3384 if (expires == cfs_b->runtime_expires)
3385 cfs_b->runtime = runtime;
3386 raw_spin_unlock(&cfs_b->lock);
3390 * When a group wakes up we want to make sure that its quota is not already
3391 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3392 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3394 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3396 if (!cfs_bandwidth_used())
3399 /* an active group must be handled by the update_curr()->put() path */
3400 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3403 /* ensure the group is not already throttled */
3404 if (cfs_rq_throttled(cfs_rq))
3407 /* update runtime allocation */
3408 account_cfs_rq_runtime(cfs_rq, 0);
3409 if (cfs_rq->runtime_remaining <= 0)
3410 throttle_cfs_rq(cfs_rq);
3413 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3414 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3416 if (!cfs_bandwidth_used())
3419 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3423 * it's possible for a throttled entity to be forced into a running
3424 * state (e.g. set_curr_task), in this case we're finished.
3426 if (cfs_rq_throttled(cfs_rq))
3429 throttle_cfs_rq(cfs_rq);
3432 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3434 struct cfs_bandwidth *cfs_b =
3435 container_of(timer, struct cfs_bandwidth, slack_timer);
3436 do_sched_cfs_slack_timer(cfs_b);
3438 return HRTIMER_NORESTART;
3441 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3443 struct cfs_bandwidth *cfs_b =
3444 container_of(timer, struct cfs_bandwidth, period_timer);
3450 now = hrtimer_cb_get_time(timer);
3451 overrun = hrtimer_forward(timer, now, cfs_b->period);
3456 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3459 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3462 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3464 raw_spin_lock_init(&cfs_b->lock);
3466 cfs_b->quota = RUNTIME_INF;
3467 cfs_b->period = ns_to_ktime(default_cfs_period());
3469 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3470 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3471 cfs_b->period_timer.function = sched_cfs_period_timer;
3472 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3473 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3476 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3478 cfs_rq->runtime_enabled = 0;
3479 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3482 /* requires cfs_b->lock, may release to reprogram timer */
3483 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3486 * The timer may be active because we're trying to set a new bandwidth
3487 * period or because we're racing with the tear-down path
3488 * (timer_active==0 becomes visible before the hrtimer call-back
3489 * terminates). In either case we ensure that it's re-programmed
3491 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3492 raw_spin_unlock(&cfs_b->lock);
3493 /* ensure cfs_b->lock is available while we wait */
3494 hrtimer_cancel(&cfs_b->period_timer);
3496 raw_spin_lock(&cfs_b->lock);
3497 /* if someone else restarted the timer then we're done */
3498 if (cfs_b->timer_active)
3502 cfs_b->timer_active = 1;
3503 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3506 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3508 hrtimer_cancel(&cfs_b->period_timer);
3509 hrtimer_cancel(&cfs_b->slack_timer);
3512 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3514 struct cfs_rq *cfs_rq;
3516 for_each_leaf_cfs_rq(rq, cfs_rq) {
3517 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3519 if (!cfs_rq->runtime_enabled)
3523 * clock_task is not advancing so we just need to make sure
3524 * there's some valid quota amount
3526 cfs_rq->runtime_remaining = cfs_b->quota;
3527 if (cfs_rq_throttled(cfs_rq))
3528 unthrottle_cfs_rq(cfs_rq);
3532 #else /* CONFIG_CFS_BANDWIDTH */
3533 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3535 return rq_clock_task(rq_of(cfs_rq));
3538 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3539 unsigned long delta_exec) {}
3540 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3541 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3542 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3544 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3549 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3554 static inline int throttled_lb_pair(struct task_group *tg,
3555 int src_cpu, int dest_cpu)
3560 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3562 #ifdef CONFIG_FAIR_GROUP_SCHED
3563 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3566 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3570 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3571 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3573 #endif /* CONFIG_CFS_BANDWIDTH */
3575 /**************************************************
3576 * CFS operations on tasks:
3579 #ifdef CONFIG_SCHED_HRTICK
3580 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3582 struct sched_entity *se = &p->se;
3583 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3585 WARN_ON(task_rq(p) != rq);
3587 if (cfs_rq->nr_running > 1) {
3588 u64 slice = sched_slice(cfs_rq, se);
3589 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3590 s64 delta = slice - ran;
3599 * Don't schedule slices shorter than 10000ns, that just
3600 * doesn't make sense. Rely on vruntime for fairness.
3603 delta = max_t(s64, 10000LL, delta);
3605 hrtick_start(rq, delta);
3610 * called from enqueue/dequeue and updates the hrtick when the
3611 * current task is from our class and nr_running is low enough
3614 static void hrtick_update(struct rq *rq)
3616 struct task_struct *curr = rq->curr;
3618 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3621 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3622 hrtick_start_fair(rq, curr);
3624 #else /* !CONFIG_SCHED_HRTICK */
3626 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3630 static inline void hrtick_update(struct rq *rq)
3636 * The enqueue_task method is called before nr_running is
3637 * increased. Here we update the fair scheduling stats and
3638 * then put the task into the rbtree:
3641 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3643 struct cfs_rq *cfs_rq;
3644 struct sched_entity *se = &p->se;
3646 for_each_sched_entity(se) {
3649 cfs_rq = cfs_rq_of(se);
3650 enqueue_entity(cfs_rq, se, flags);
3653 * end evaluation on encountering a throttled cfs_rq
3655 * note: in the case of encountering a throttled cfs_rq we will
3656 * post the final h_nr_running increment below.
3658 if (cfs_rq_throttled(cfs_rq))
3660 cfs_rq->h_nr_running++;
3662 flags = ENQUEUE_WAKEUP;
3665 for_each_sched_entity(se) {
3666 cfs_rq = cfs_rq_of(se);
3667 cfs_rq->h_nr_running++;
3669 if (cfs_rq_throttled(cfs_rq))
3672 update_cfs_shares(cfs_rq);
3673 update_entity_load_avg(se, 1);
3677 update_rq_runnable_avg(rq, rq->nr_running);
3683 static void set_next_buddy(struct sched_entity *se);
3686 * The dequeue_task method is called before nr_running is
3687 * decreased. We remove the task from the rbtree and
3688 * update the fair scheduling stats:
3690 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3692 struct cfs_rq *cfs_rq;
3693 struct sched_entity *se = &p->se;
3694 int task_sleep = flags & DEQUEUE_SLEEP;
3696 for_each_sched_entity(se) {
3697 cfs_rq = cfs_rq_of(se);
3698 dequeue_entity(cfs_rq, se, flags);
3701 * end evaluation on encountering a throttled cfs_rq
3703 * note: in the case of encountering a throttled cfs_rq we will
3704 * post the final h_nr_running decrement below.
3706 if (cfs_rq_throttled(cfs_rq))
3708 cfs_rq->h_nr_running--;
3710 /* Don't dequeue parent if it has other entities besides us */
3711 if (cfs_rq->load.weight) {
3713 * Bias pick_next to pick a task from this cfs_rq, as
3714 * p is sleeping when it is within its sched_slice.
3716 if (task_sleep && parent_entity(se))
3717 set_next_buddy(parent_entity(se));
3719 /* avoid re-evaluating load for this entity */
3720 se = parent_entity(se);
3723 flags |= DEQUEUE_SLEEP;
3726 for_each_sched_entity(se) {
3727 cfs_rq = cfs_rq_of(se);
3728 cfs_rq->h_nr_running--;
3730 if (cfs_rq_throttled(cfs_rq))
3733 update_cfs_shares(cfs_rq);
3734 update_entity_load_avg(se, 1);
3739 update_rq_runnable_avg(rq, 1);
3745 /* Used instead of source_load when we know the type == 0 */
3746 static unsigned long weighted_cpuload(const int cpu)
3748 return cpu_rq(cpu)->cfs.runnable_load_avg;
3752 * Return a low guess at the load of a migration-source cpu weighted
3753 * according to the scheduling class and "nice" value.
3755 * We want to under-estimate the load of migration sources, to
3756 * balance conservatively.
3758 static unsigned long source_load(int cpu, int type)
3760 struct rq *rq = cpu_rq(cpu);
3761 unsigned long total = weighted_cpuload(cpu);
3763 if (type == 0 || !sched_feat(LB_BIAS))
3766 return min(rq->cpu_load[type-1], total);
3770 * Return a high guess at the load of a migration-target cpu weighted
3771 * according to the scheduling class and "nice" value.
3773 static unsigned long target_load(int cpu, int type)
3775 struct rq *rq = cpu_rq(cpu);
3776 unsigned long total = weighted_cpuload(cpu);
3778 if (type == 0 || !sched_feat(LB_BIAS))
3781 return max(rq->cpu_load[type-1], total);
3784 static unsigned long power_of(int cpu)
3786 return cpu_rq(cpu)->cpu_power;
3789 static unsigned long cpu_avg_load_per_task(int cpu)
3791 struct rq *rq = cpu_rq(cpu);
3792 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3793 unsigned long load_avg = rq->cfs.runnable_load_avg;
3796 return load_avg / nr_running;
3801 static void record_wakee(struct task_struct *p)
3804 * Rough decay (wiping) for cost saving, don't worry
3805 * about the boundary, really active task won't care
3808 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3809 current->wakee_flips = 0;
3810 current->wakee_flip_decay_ts = jiffies;
3813 if (current->last_wakee != p) {
3814 current->last_wakee = p;
3815 current->wakee_flips++;
3819 static void task_waking_fair(struct task_struct *p)
3821 struct sched_entity *se = &p->se;
3822 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3825 #ifndef CONFIG_64BIT
3826 u64 min_vruntime_copy;
3829 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3831 min_vruntime = cfs_rq->min_vruntime;
3832 } while (min_vruntime != min_vruntime_copy);
3834 min_vruntime = cfs_rq->min_vruntime;
3837 se->vruntime -= min_vruntime;
3841 #ifdef CONFIG_FAIR_GROUP_SCHED
3843 * effective_load() calculates the load change as seen from the root_task_group
3845 * Adding load to a group doesn't make a group heavier, but can cause movement
3846 * of group shares between cpus. Assuming the shares were perfectly aligned one
3847 * can calculate the shift in shares.
3849 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3850 * on this @cpu and results in a total addition (subtraction) of @wg to the
3851 * total group weight.
3853 * Given a runqueue weight distribution (rw_i) we can compute a shares
3854 * distribution (s_i) using:
3856 * s_i = rw_i / \Sum rw_j (1)
3858 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3859 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3860 * shares distribution (s_i):
3862 * rw_i = { 2, 4, 1, 0 }
3863 * s_i = { 2/7, 4/7, 1/7, 0 }
3865 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3866 * task used to run on and the CPU the waker is running on), we need to
3867 * compute the effect of waking a task on either CPU and, in case of a sync
3868 * wakeup, compute the effect of the current task going to sleep.
3870 * So for a change of @wl to the local @cpu with an overall group weight change
3871 * of @wl we can compute the new shares distribution (s'_i) using:
3873 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3875 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3876 * differences in waking a task to CPU 0. The additional task changes the
3877 * weight and shares distributions like:
3879 * rw'_i = { 3, 4, 1, 0 }
3880 * s'_i = { 3/8, 4/8, 1/8, 0 }
3882 * We can then compute the difference in effective weight by using:
3884 * dw_i = S * (s'_i - s_i) (3)
3886 * Where 'S' is the group weight as seen by its parent.
3888 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3889 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3890 * 4/7) times the weight of the group.
3892 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3894 struct sched_entity *se = tg->se[cpu];
3896 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3899 for_each_sched_entity(se) {
3905 * W = @wg + \Sum rw_j
3907 W = wg + calc_tg_weight(tg, se->my_q);
3912 w = se->my_q->load.weight + wl;
3915 * wl = S * s'_i; see (2)
3918 wl = (w * tg->shares) / W;
3923 * Per the above, wl is the new se->load.weight value; since
3924 * those are clipped to [MIN_SHARES, ...) do so now. See
3925 * calc_cfs_shares().
3927 if (wl < MIN_SHARES)
3931 * wl = dw_i = S * (s'_i - s_i); see (3)
3933 wl -= se->load.weight;
3936 * Recursively apply this logic to all parent groups to compute
3937 * the final effective load change on the root group. Since
3938 * only the @tg group gets extra weight, all parent groups can
3939 * only redistribute existing shares. @wl is the shift in shares
3940 * resulting from this level per the above.
3949 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3956 static int wake_wide(struct task_struct *p)
3958 int factor = this_cpu_read(sd_llc_size);
3961 * Yeah, it's the switching-frequency, could means many wakee or
3962 * rapidly switch, use factor here will just help to automatically
3963 * adjust the loose-degree, so bigger node will lead to more pull.
3965 if (p->wakee_flips > factor) {
3967 * wakee is somewhat hot, it needs certain amount of cpu
3968 * resource, so if waker is far more hot, prefer to leave
3971 if (current->wakee_flips > (factor * p->wakee_flips))
3978 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3980 s64 this_load, load;
3981 int idx, this_cpu, prev_cpu;
3982 unsigned long tl_per_task;
3983 struct task_group *tg;
3984 unsigned long weight;
3988 * If we wake multiple tasks be careful to not bounce
3989 * ourselves around too much.
3995 this_cpu = smp_processor_id();
3996 prev_cpu = task_cpu(p);
3997 load = source_load(prev_cpu, idx);
3998 this_load = target_load(this_cpu, idx);
4001 * If sync wakeup then subtract the (maximum possible)
4002 * effect of the currently running task from the load
4003 * of the current CPU:
4006 tg = task_group(current);
4007 weight = current->se.load.weight;
4009 this_load += effective_load(tg, this_cpu, -weight, -weight);
4010 load += effective_load(tg, prev_cpu, 0, -weight);
4014 weight = p->se.load.weight;
4017 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4018 * due to the sync cause above having dropped this_load to 0, we'll
4019 * always have an imbalance, but there's really nothing you can do
4020 * about that, so that's good too.
4022 * Otherwise check if either cpus are near enough in load to allow this
4023 * task to be woken on this_cpu.
4025 if (this_load > 0) {
4026 s64 this_eff_load, prev_eff_load;
4028 this_eff_load = 100;
4029 this_eff_load *= power_of(prev_cpu);
4030 this_eff_load *= this_load +
4031 effective_load(tg, this_cpu, weight, weight);
4033 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4034 prev_eff_load *= power_of(this_cpu);
4035 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4037 balanced = this_eff_load <= prev_eff_load;
4042 * If the currently running task will sleep within
4043 * a reasonable amount of time then attract this newly
4046 if (sync && balanced)
4049 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4050 tl_per_task = cpu_avg_load_per_task(this_cpu);
4053 (this_load <= load &&
4054 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4056 * This domain has SD_WAKE_AFFINE and
4057 * p is cache cold in this domain, and
4058 * there is no bad imbalance.
4060 schedstat_inc(sd, ttwu_move_affine);
4061 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4069 * find_idlest_group finds and returns the least busy CPU group within the
4072 static struct sched_group *
4073 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4074 int this_cpu, int load_idx)
4076 struct sched_group *idlest = NULL, *group = sd->groups;
4077 unsigned long min_load = ULONG_MAX, this_load = 0;
4078 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4081 unsigned long load, avg_load;
4085 /* Skip over this group if it has no CPUs allowed */
4086 if (!cpumask_intersects(sched_group_cpus(group),
4087 tsk_cpus_allowed(p)))
4090 local_group = cpumask_test_cpu(this_cpu,
4091 sched_group_cpus(group));
4093 /* Tally up the load of all CPUs in the group */
4096 for_each_cpu(i, sched_group_cpus(group)) {
4097 /* Bias balancing toward cpus of our domain */
4099 load = source_load(i, load_idx);
4101 load = target_load(i, load_idx);
4106 /* Adjust by relative CPU power of the group */
4107 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4110 this_load = avg_load;
4111 } else if (avg_load < min_load) {
4112 min_load = avg_load;
4115 } while (group = group->next, group != sd->groups);
4117 if (!idlest || 100*this_load < imbalance*min_load)
4123 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4126 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4128 unsigned long load, min_load = ULONG_MAX;
4132 /* Traverse only the allowed CPUs */
4133 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4134 load = weighted_cpuload(i);
4136 if (load < min_load || (load == min_load && i == this_cpu)) {
4146 * Try and locate an idle CPU in the sched_domain.
4148 static int select_idle_sibling(struct task_struct *p, int target)
4150 struct sched_domain *sd;
4151 struct sched_group *sg;
4152 int i = task_cpu(p);
4154 if (idle_cpu(target))
4158 * If the prevous cpu is cache affine and idle, don't be stupid.
4160 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4164 * Otherwise, iterate the domains and find an elegible idle cpu.
4166 sd = rcu_dereference(per_cpu(sd_llc, target));
4167 for_each_lower_domain(sd) {
4170 if (!cpumask_intersects(sched_group_cpus(sg),
4171 tsk_cpus_allowed(p)))
4174 for_each_cpu(i, sched_group_cpus(sg)) {
4175 if (i == target || !idle_cpu(i))
4179 target = cpumask_first_and(sched_group_cpus(sg),
4180 tsk_cpus_allowed(p));
4184 } while (sg != sd->groups);
4191 * sched_balance_self: balance the current task (running on cpu) in domains
4192 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4195 * Balance, ie. select the least loaded group.
4197 * Returns the target CPU number, or the same CPU if no balancing is needed.
4199 * preempt must be disabled.
4202 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4204 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4205 int cpu = smp_processor_id();
4207 int want_affine = 0;
4208 int sync = wake_flags & WF_SYNC;
4210 if (p->nr_cpus_allowed == 1)
4213 if (sd_flag & SD_BALANCE_WAKE) {
4214 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4220 for_each_domain(cpu, tmp) {
4221 if (!(tmp->flags & SD_LOAD_BALANCE))
4225 * If both cpu and prev_cpu are part of this domain,
4226 * cpu is a valid SD_WAKE_AFFINE target.
4228 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4229 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4234 if (tmp->flags & sd_flag)
4239 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4242 new_cpu = select_idle_sibling(p, prev_cpu);
4247 int load_idx = sd->forkexec_idx;
4248 struct sched_group *group;
4251 if (!(sd->flags & sd_flag)) {
4256 if (sd_flag & SD_BALANCE_WAKE)
4257 load_idx = sd->wake_idx;
4259 group = find_idlest_group(sd, p, cpu, load_idx);
4265 new_cpu = find_idlest_cpu(group, p, cpu);
4266 if (new_cpu == -1 || new_cpu == cpu) {
4267 /* Now try balancing at a lower domain level of cpu */
4272 /* Now try balancing at a lower domain level of new_cpu */
4274 weight = sd->span_weight;
4276 for_each_domain(cpu, tmp) {
4277 if (weight <= tmp->span_weight)
4279 if (tmp->flags & sd_flag)
4282 /* while loop will break here if sd == NULL */
4291 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4292 * cfs_rq_of(p) references at time of call are still valid and identify the
4293 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4294 * other assumptions, including the state of rq->lock, should be made.
4297 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4299 struct sched_entity *se = &p->se;
4300 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4303 * Load tracking: accumulate removed load so that it can be processed
4304 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4305 * to blocked load iff they have a positive decay-count. It can never
4306 * be negative here since on-rq tasks have decay-count == 0.
4308 if (se->avg.decay_count) {
4309 se->avg.decay_count = -__synchronize_entity_decay(se);
4310 atomic_long_add(se->avg.load_avg_contrib,
4311 &cfs_rq->removed_load);
4314 #endif /* CONFIG_SMP */
4316 static unsigned long
4317 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4319 unsigned long gran = sysctl_sched_wakeup_granularity;
4322 * Since its curr running now, convert the gran from real-time
4323 * to virtual-time in his units.
4325 * By using 'se' instead of 'curr' we penalize light tasks, so
4326 * they get preempted easier. That is, if 'se' < 'curr' then
4327 * the resulting gran will be larger, therefore penalizing the
4328 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4329 * be smaller, again penalizing the lighter task.
4331 * This is especially important for buddies when the leftmost
4332 * task is higher priority than the buddy.
4334 return calc_delta_fair(gran, se);
4338 * Should 'se' preempt 'curr'.
4352 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4354 s64 gran, vdiff = curr->vruntime - se->vruntime;
4359 gran = wakeup_gran(curr, se);
4366 static void set_last_buddy(struct sched_entity *se)
4368 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4371 for_each_sched_entity(se)
4372 cfs_rq_of(se)->last = se;
4375 static void set_next_buddy(struct sched_entity *se)
4377 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4380 for_each_sched_entity(se)
4381 cfs_rq_of(se)->next = se;
4384 static void set_skip_buddy(struct sched_entity *se)
4386 for_each_sched_entity(se)
4387 cfs_rq_of(se)->skip = se;
4391 * Preempt the current task with a newly woken task if needed:
4393 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4395 struct task_struct *curr = rq->curr;
4396 struct sched_entity *se = &curr->se, *pse = &p->se;
4397 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4398 int scale = cfs_rq->nr_running >= sched_nr_latency;
4399 int next_buddy_marked = 0;
4401 if (unlikely(se == pse))
4405 * This is possible from callers such as move_task(), in which we
4406 * unconditionally check_prempt_curr() after an enqueue (which may have
4407 * lead to a throttle). This both saves work and prevents false
4408 * next-buddy nomination below.
4410 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4413 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4414 set_next_buddy(pse);
4415 next_buddy_marked = 1;
4419 * We can come here with TIF_NEED_RESCHED already set from new task
4422 * Note: this also catches the edge-case of curr being in a throttled
4423 * group (e.g. via set_curr_task), since update_curr() (in the
4424 * enqueue of curr) will have resulted in resched being set. This
4425 * prevents us from potentially nominating it as a false LAST_BUDDY
4428 if (test_tsk_need_resched(curr))
4431 /* Idle tasks are by definition preempted by non-idle tasks. */
4432 if (unlikely(curr->policy == SCHED_IDLE) &&
4433 likely(p->policy != SCHED_IDLE))
4437 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4438 * is driven by the tick):
4440 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4443 find_matching_se(&se, &pse);
4444 update_curr(cfs_rq_of(se));
4446 if (wakeup_preempt_entity(se, pse) == 1) {
4448 * Bias pick_next to pick the sched entity that is
4449 * triggering this preemption.
4451 if (!next_buddy_marked)
4452 set_next_buddy(pse);
4461 * Only set the backward buddy when the current task is still
4462 * on the rq. This can happen when a wakeup gets interleaved
4463 * with schedule on the ->pre_schedule() or idle_balance()
4464 * point, either of which can * drop the rq lock.
4466 * Also, during early boot the idle thread is in the fair class,
4467 * for obvious reasons its a bad idea to schedule back to it.
4469 if (unlikely(!se->on_rq || curr == rq->idle))
4472 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4476 static struct task_struct *pick_next_task_fair(struct rq *rq)
4478 struct task_struct *p;
4479 struct cfs_rq *cfs_rq = &rq->cfs;
4480 struct sched_entity *se;
4482 if (!cfs_rq->nr_running)
4486 se = pick_next_entity(cfs_rq);
4487 set_next_entity(cfs_rq, se);
4488 cfs_rq = group_cfs_rq(se);
4492 if (hrtick_enabled(rq))
4493 hrtick_start_fair(rq, p);
4499 * Account for a descheduled task:
4501 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4503 struct sched_entity *se = &prev->se;
4504 struct cfs_rq *cfs_rq;
4506 for_each_sched_entity(se) {
4507 cfs_rq = cfs_rq_of(se);
4508 put_prev_entity(cfs_rq, se);
4513 * sched_yield() is very simple
4515 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4517 static void yield_task_fair(struct rq *rq)
4519 struct task_struct *curr = rq->curr;
4520 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4521 struct sched_entity *se = &curr->se;
4524 * Are we the only task in the tree?
4526 if (unlikely(rq->nr_running == 1))
4529 clear_buddies(cfs_rq, se);
4531 if (curr->policy != SCHED_BATCH) {
4532 update_rq_clock(rq);
4534 * Update run-time statistics of the 'current'.
4536 update_curr(cfs_rq);
4538 * Tell update_rq_clock() that we've just updated,
4539 * so we don't do microscopic update in schedule()
4540 * and double the fastpath cost.
4542 rq->skip_clock_update = 1;
4548 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4550 struct sched_entity *se = &p->se;
4552 /* throttled hierarchies are not runnable */
4553 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4556 /* Tell the scheduler that we'd really like pse to run next. */
4559 yield_task_fair(rq);
4565 /**************************************************
4566 * Fair scheduling class load-balancing methods.
4570 * The purpose of load-balancing is to achieve the same basic fairness the
4571 * per-cpu scheduler provides, namely provide a proportional amount of compute
4572 * time to each task. This is expressed in the following equation:
4574 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4576 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4577 * W_i,0 is defined as:
4579 * W_i,0 = \Sum_j w_i,j (2)
4581 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4582 * is derived from the nice value as per prio_to_weight[].
4584 * The weight average is an exponential decay average of the instantaneous
4587 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4589 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4590 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4591 * can also include other factors [XXX].
4593 * To achieve this balance we define a measure of imbalance which follows
4594 * directly from (1):
4596 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4598 * We them move tasks around to minimize the imbalance. In the continuous
4599 * function space it is obvious this converges, in the discrete case we get
4600 * a few fun cases generally called infeasible weight scenarios.
4603 * - infeasible weights;
4604 * - local vs global optima in the discrete case. ]
4609 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4610 * for all i,j solution, we create a tree of cpus that follows the hardware
4611 * topology where each level pairs two lower groups (or better). This results
4612 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4613 * tree to only the first of the previous level and we decrease the frequency
4614 * of load-balance at each level inv. proportional to the number of cpus in
4620 * \Sum { --- * --- * 2^i } = O(n) (5)
4622 * `- size of each group
4623 * | | `- number of cpus doing load-balance
4625 * `- sum over all levels
4627 * Coupled with a limit on how many tasks we can migrate every balance pass,
4628 * this makes (5) the runtime complexity of the balancer.
4630 * An important property here is that each CPU is still (indirectly) connected
4631 * to every other cpu in at most O(log n) steps:
4633 * The adjacency matrix of the resulting graph is given by:
4636 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4639 * And you'll find that:
4641 * A^(log_2 n)_i,j != 0 for all i,j (7)
4643 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4644 * The task movement gives a factor of O(m), giving a convergence complexity
4647 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4652 * In order to avoid CPUs going idle while there's still work to do, new idle
4653 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4654 * tree itself instead of relying on other CPUs to bring it work.
4656 * This adds some complexity to both (5) and (8) but it reduces the total idle
4664 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4667 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4672 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4674 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4676 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4679 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4680 * rewrite all of this once again.]
4683 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4685 enum fbq_type { regular, remote, all };
4687 #define LBF_ALL_PINNED 0x01
4688 #define LBF_NEED_BREAK 0x02
4689 #define LBF_DST_PINNED 0x04
4690 #define LBF_SOME_PINNED 0x08
4693 struct sched_domain *sd;
4701 struct cpumask *dst_grpmask;
4703 enum cpu_idle_type idle;
4705 /* The set of CPUs under consideration for load-balancing */
4706 struct cpumask *cpus;
4711 unsigned int loop_break;
4712 unsigned int loop_max;
4714 enum fbq_type fbq_type;
4718 * move_task - move a task from one runqueue to another runqueue.
4719 * Both runqueues must be locked.
4721 static void move_task(struct task_struct *p, struct lb_env *env)
4723 deactivate_task(env->src_rq, p, 0);
4724 set_task_cpu(p, env->dst_cpu);
4725 activate_task(env->dst_rq, p, 0);
4726 check_preempt_curr(env->dst_rq, p, 0);
4730 * Is this task likely cache-hot:
4733 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4737 if (p->sched_class != &fair_sched_class)
4740 if (unlikely(p->policy == SCHED_IDLE))
4744 * Buddy candidates are cache hot:
4746 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4747 (&p->se == cfs_rq_of(&p->se)->next ||
4748 &p->se == cfs_rq_of(&p->se)->last))
4751 if (sysctl_sched_migration_cost == -1)
4753 if (sysctl_sched_migration_cost == 0)
4756 delta = now - p->se.exec_start;
4758 return delta < (s64)sysctl_sched_migration_cost;
4761 #ifdef CONFIG_NUMA_BALANCING
4762 /* Returns true if the destination node has incurred more faults */
4763 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4765 int src_nid, dst_nid;
4767 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4768 !(env->sd->flags & SD_NUMA)) {
4772 src_nid = cpu_to_node(env->src_cpu);
4773 dst_nid = cpu_to_node(env->dst_cpu);
4775 if (src_nid == dst_nid)
4778 /* Always encourage migration to the preferred node. */
4779 if (dst_nid == p->numa_preferred_nid)
4782 /* If both task and group weight improve, this move is a winner. */
4783 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4784 group_weight(p, dst_nid) > group_weight(p, src_nid))
4791 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4793 int src_nid, dst_nid;
4795 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4798 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4801 src_nid = cpu_to_node(env->src_cpu);
4802 dst_nid = cpu_to_node(env->dst_cpu);
4804 if (src_nid == dst_nid)
4807 /* Migrating away from the preferred node is always bad. */
4808 if (src_nid == p->numa_preferred_nid)
4811 /* If either task or group weight get worse, don't do it. */
4812 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4813 group_weight(p, dst_nid) < group_weight(p, src_nid))
4820 static inline bool migrate_improves_locality(struct task_struct *p,
4826 static inline bool migrate_degrades_locality(struct task_struct *p,
4834 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4837 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4839 int tsk_cache_hot = 0;
4841 * We do not migrate tasks that are:
4842 * 1) throttled_lb_pair, or
4843 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4844 * 3) running (obviously), or
4845 * 4) are cache-hot on their current CPU.
4847 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4850 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4853 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4855 env->flags |= LBF_SOME_PINNED;
4858 * Remember if this task can be migrated to any other cpu in
4859 * our sched_group. We may want to revisit it if we couldn't
4860 * meet load balance goals by pulling other tasks on src_cpu.
4862 * Also avoid computing new_dst_cpu if we have already computed
4863 * one in current iteration.
4865 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4868 /* Prevent to re-select dst_cpu via env's cpus */
4869 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4870 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4871 env->flags |= LBF_DST_PINNED;
4872 env->new_dst_cpu = cpu;
4880 /* Record that we found atleast one task that could run on dst_cpu */
4881 env->flags &= ~LBF_ALL_PINNED;
4883 if (task_running(env->src_rq, p)) {
4884 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4889 * Aggressive migration if:
4890 * 1) destination numa is preferred
4891 * 2) task is cache cold, or
4892 * 3) too many balance attempts have failed.
4894 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4896 tsk_cache_hot = migrate_degrades_locality(p, env);
4898 if (migrate_improves_locality(p, env)) {
4899 #ifdef CONFIG_SCHEDSTATS
4900 if (tsk_cache_hot) {
4901 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4902 schedstat_inc(p, se.statistics.nr_forced_migrations);
4908 if (!tsk_cache_hot ||
4909 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4911 if (tsk_cache_hot) {
4912 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4913 schedstat_inc(p, se.statistics.nr_forced_migrations);
4919 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4924 * move_one_task tries to move exactly one task from busiest to this_rq, as
4925 * part of active balancing operations within "domain".
4926 * Returns 1 if successful and 0 otherwise.
4928 * Called with both runqueues locked.
4930 static int move_one_task(struct lb_env *env)
4932 struct task_struct *p, *n;
4934 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4935 if (!can_migrate_task(p, env))
4940 * Right now, this is only the second place move_task()
4941 * is called, so we can safely collect move_task()
4942 * stats here rather than inside move_task().
4944 schedstat_inc(env->sd, lb_gained[env->idle]);
4950 static const unsigned int sched_nr_migrate_break = 32;
4953 * move_tasks tries to move up to imbalance weighted load from busiest to
4954 * this_rq, as part of a balancing operation within domain "sd".
4955 * Returns 1 if successful and 0 otherwise.
4957 * Called with both runqueues locked.
4959 static int move_tasks(struct lb_env *env)
4961 struct list_head *tasks = &env->src_rq->cfs_tasks;
4962 struct task_struct *p;
4966 if (env->imbalance <= 0)
4969 while (!list_empty(tasks)) {
4970 p = list_first_entry(tasks, struct task_struct, se.group_node);
4973 /* We've more or less seen every task there is, call it quits */
4974 if (env->loop > env->loop_max)
4977 /* take a breather every nr_migrate tasks */
4978 if (env->loop > env->loop_break) {
4979 env->loop_break += sched_nr_migrate_break;
4980 env->flags |= LBF_NEED_BREAK;
4984 if (!can_migrate_task(p, env))
4987 load = task_h_load(p);
4989 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4992 if ((load / 2) > env->imbalance)
4997 env->imbalance -= load;
4999 #ifdef CONFIG_PREEMPT
5001 * NEWIDLE balancing is a source of latency, so preemptible
5002 * kernels will stop after the first task is pulled to minimize
5003 * the critical section.
5005 if (env->idle == CPU_NEWLY_IDLE)
5010 * We only want to steal up to the prescribed amount of
5013 if (env->imbalance <= 0)
5018 list_move_tail(&p->se.group_node, tasks);
5022 * Right now, this is one of only two places move_task() is called,
5023 * so we can safely collect move_task() stats here rather than
5024 * inside move_task().
5026 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5031 #ifdef CONFIG_FAIR_GROUP_SCHED
5033 * update tg->load_weight by folding this cpu's load_avg
5035 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5037 struct sched_entity *se = tg->se[cpu];
5038 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5040 /* throttled entities do not contribute to load */
5041 if (throttled_hierarchy(cfs_rq))
5044 update_cfs_rq_blocked_load(cfs_rq, 1);
5047 update_entity_load_avg(se, 1);
5049 * We pivot on our runnable average having decayed to zero for
5050 * list removal. This generally implies that all our children
5051 * have also been removed (modulo rounding error or bandwidth
5052 * control); however, such cases are rare and we can fix these
5055 * TODO: fix up out-of-order children on enqueue.
5057 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5058 list_del_leaf_cfs_rq(cfs_rq);
5060 struct rq *rq = rq_of(cfs_rq);
5061 update_rq_runnable_avg(rq, rq->nr_running);
5065 static void update_blocked_averages(int cpu)
5067 struct rq *rq = cpu_rq(cpu);
5068 struct cfs_rq *cfs_rq;
5069 unsigned long flags;
5071 raw_spin_lock_irqsave(&rq->lock, flags);
5072 update_rq_clock(rq);
5074 * Iterates the task_group tree in a bottom up fashion, see
5075 * list_add_leaf_cfs_rq() for details.
5077 for_each_leaf_cfs_rq(rq, cfs_rq) {
5079 * Note: We may want to consider periodically releasing
5080 * rq->lock about these updates so that creating many task
5081 * groups does not result in continually extending hold time.
5083 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5086 raw_spin_unlock_irqrestore(&rq->lock, flags);
5090 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5091 * This needs to be done in a top-down fashion because the load of a child
5092 * group is a fraction of its parents load.
5094 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5096 struct rq *rq = rq_of(cfs_rq);
5097 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5098 unsigned long now = jiffies;
5101 if (cfs_rq->last_h_load_update == now)
5104 cfs_rq->h_load_next = NULL;
5105 for_each_sched_entity(se) {
5106 cfs_rq = cfs_rq_of(se);
5107 cfs_rq->h_load_next = se;
5108 if (cfs_rq->last_h_load_update == now)
5113 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5114 cfs_rq->last_h_load_update = now;
5117 while ((se = cfs_rq->h_load_next) != NULL) {
5118 load = cfs_rq->h_load;
5119 load = div64_ul(load * se->avg.load_avg_contrib,
5120 cfs_rq->runnable_load_avg + 1);
5121 cfs_rq = group_cfs_rq(se);
5122 cfs_rq->h_load = load;
5123 cfs_rq->last_h_load_update = now;
5127 static unsigned long task_h_load(struct task_struct *p)
5129 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5131 update_cfs_rq_h_load(cfs_rq);
5132 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5133 cfs_rq->runnable_load_avg + 1);
5136 static inline void update_blocked_averages(int cpu)
5140 static unsigned long task_h_load(struct task_struct *p)
5142 return p->se.avg.load_avg_contrib;
5146 /********** Helpers for find_busiest_group ************************/
5148 * sg_lb_stats - stats of a sched_group required for load_balancing
5150 struct sg_lb_stats {
5151 unsigned long avg_load; /*Avg load across the CPUs of the group */
5152 unsigned long group_load; /* Total load over the CPUs of the group */
5153 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5154 unsigned long load_per_task;
5155 unsigned long group_power;
5156 unsigned int sum_nr_running; /* Nr tasks running in the group */
5157 unsigned int group_capacity;
5158 unsigned int idle_cpus;
5159 unsigned int group_weight;
5160 int group_imb; /* Is there an imbalance in the group ? */
5161 int group_has_capacity; /* Is there extra capacity in the group? */
5162 #ifdef CONFIG_NUMA_BALANCING
5163 unsigned int nr_numa_running;
5164 unsigned int nr_preferred_running;
5169 * sd_lb_stats - Structure to store the statistics of a sched_domain
5170 * during load balancing.
5172 struct sd_lb_stats {
5173 struct sched_group *busiest; /* Busiest group in this sd */
5174 struct sched_group *local; /* Local group in this sd */
5175 unsigned long total_load; /* Total load of all groups in sd */
5176 unsigned long total_pwr; /* Total power of all groups in sd */
5177 unsigned long avg_load; /* Average load across all groups in sd */
5179 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5180 struct sg_lb_stats local_stat; /* Statistics of the local group */
5183 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5186 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5187 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5188 * We must however clear busiest_stat::avg_load because
5189 * update_sd_pick_busiest() reads this before assignment.
5191 *sds = (struct sd_lb_stats){
5203 * get_sd_load_idx - Obtain the load index for a given sched domain.
5204 * @sd: The sched_domain whose load_idx is to be obtained.
5205 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5207 * Return: The load index.
5209 static inline int get_sd_load_idx(struct sched_domain *sd,
5210 enum cpu_idle_type idle)
5216 load_idx = sd->busy_idx;
5219 case CPU_NEWLY_IDLE:
5220 load_idx = sd->newidle_idx;
5223 load_idx = sd->idle_idx;
5230 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5232 return SCHED_POWER_SCALE;
5235 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5237 return default_scale_freq_power(sd, cpu);
5240 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5242 unsigned long weight = sd->span_weight;
5243 unsigned long smt_gain = sd->smt_gain;
5250 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5252 return default_scale_smt_power(sd, cpu);
5255 static unsigned long scale_rt_power(int cpu)
5257 struct rq *rq = cpu_rq(cpu);
5258 u64 total, available, age_stamp, avg;
5261 * Since we're reading these variables without serialization make sure
5262 * we read them once before doing sanity checks on them.
5264 age_stamp = ACCESS_ONCE(rq->age_stamp);
5265 avg = ACCESS_ONCE(rq->rt_avg);
5267 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5269 if (unlikely(total < avg)) {
5270 /* Ensures that power won't end up being negative */
5273 available = total - avg;
5276 if (unlikely((s64)total < SCHED_POWER_SCALE))
5277 total = SCHED_POWER_SCALE;
5279 total >>= SCHED_POWER_SHIFT;
5281 return div_u64(available, total);
5284 static void update_cpu_power(struct sched_domain *sd, int cpu)
5286 unsigned long weight = sd->span_weight;
5287 unsigned long power = SCHED_POWER_SCALE;
5288 struct sched_group *sdg = sd->groups;
5290 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5291 if (sched_feat(ARCH_POWER))
5292 power *= arch_scale_smt_power(sd, cpu);
5294 power *= default_scale_smt_power(sd, cpu);
5296 power >>= SCHED_POWER_SHIFT;
5299 sdg->sgp->power_orig = power;
5301 if (sched_feat(ARCH_POWER))
5302 power *= arch_scale_freq_power(sd, cpu);
5304 power *= default_scale_freq_power(sd, cpu);
5306 power >>= SCHED_POWER_SHIFT;
5308 power *= scale_rt_power(cpu);
5309 power >>= SCHED_POWER_SHIFT;
5314 cpu_rq(cpu)->cpu_power = power;
5315 sdg->sgp->power = power;
5318 void update_group_power(struct sched_domain *sd, int cpu)
5320 struct sched_domain *child = sd->child;
5321 struct sched_group *group, *sdg = sd->groups;
5322 unsigned long power, power_orig;
5323 unsigned long interval;
5325 interval = msecs_to_jiffies(sd->balance_interval);
5326 interval = clamp(interval, 1UL, max_load_balance_interval);
5327 sdg->sgp->next_update = jiffies + interval;
5330 update_cpu_power(sd, cpu);
5334 power_orig = power = 0;
5336 if (child->flags & SD_OVERLAP) {
5338 * SD_OVERLAP domains cannot assume that child groups
5339 * span the current group.
5342 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5343 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5345 power_orig += sg->sgp->power_orig;
5346 power += sg->sgp->power;
5350 * !SD_OVERLAP domains can assume that child groups
5351 * span the current group.
5354 group = child->groups;
5356 power_orig += group->sgp->power_orig;
5357 power += group->sgp->power;
5358 group = group->next;
5359 } while (group != child->groups);
5362 sdg->sgp->power_orig = power_orig;
5363 sdg->sgp->power = power;
5367 * Try and fix up capacity for tiny siblings, this is needed when
5368 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5369 * which on its own isn't powerful enough.
5371 * See update_sd_pick_busiest() and check_asym_packing().
5374 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5377 * Only siblings can have significantly less than SCHED_POWER_SCALE
5379 if (!(sd->flags & SD_SHARE_CPUPOWER))
5383 * If ~90% of the cpu_power is still there, we're good.
5385 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5392 * Group imbalance indicates (and tries to solve) the problem where balancing
5393 * groups is inadequate due to tsk_cpus_allowed() constraints.
5395 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5396 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5399 * { 0 1 2 3 } { 4 5 6 7 }
5402 * If we were to balance group-wise we'd place two tasks in the first group and
5403 * two tasks in the second group. Clearly this is undesired as it will overload
5404 * cpu 3 and leave one of the cpus in the second group unused.
5406 * The current solution to this issue is detecting the skew in the first group
5407 * by noticing the lower domain failed to reach balance and had difficulty
5408 * moving tasks due to affinity constraints.
5410 * When this is so detected; this group becomes a candidate for busiest; see
5411 * update_sd_pick_busiest(). And calculcate_imbalance() and
5412 * find_busiest_group() avoid some of the usual balance conditions to allow it
5413 * to create an effective group imbalance.
5415 * This is a somewhat tricky proposition since the next run might not find the
5416 * group imbalance and decide the groups need to be balanced again. A most
5417 * subtle and fragile situation.
5420 static inline int sg_imbalanced(struct sched_group *group)
5422 return group->sgp->imbalance;
5426 * Compute the group capacity.
5428 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5429 * first dividing out the smt factor and computing the actual number of cores
5430 * and limit power unit capacity with that.
5432 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5434 unsigned int capacity, smt, cpus;
5435 unsigned int power, power_orig;
5437 power = group->sgp->power;
5438 power_orig = group->sgp->power_orig;
5439 cpus = group->group_weight;
5441 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5442 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5443 capacity = cpus / smt; /* cores */
5445 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5447 capacity = fix_small_capacity(env->sd, group);
5453 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5454 * @env: The load balancing environment.
5455 * @group: sched_group whose statistics are to be updated.
5456 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5457 * @local_group: Does group contain this_cpu.
5458 * @sgs: variable to hold the statistics for this group.
5460 static inline void update_sg_lb_stats(struct lb_env *env,
5461 struct sched_group *group, int load_idx,
5462 int local_group, struct sg_lb_stats *sgs)
5464 unsigned long nr_running;
5468 memset(sgs, 0, sizeof(*sgs));
5470 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5471 struct rq *rq = cpu_rq(i);
5473 nr_running = rq->nr_running;
5475 /* Bias balancing toward cpus of our domain */
5477 load = target_load(i, load_idx);
5479 load = source_load(i, load_idx);
5481 sgs->group_load += load;
5482 sgs->sum_nr_running += nr_running;
5483 #ifdef CONFIG_NUMA_BALANCING
5484 sgs->nr_numa_running += rq->nr_numa_running;
5485 sgs->nr_preferred_running += rq->nr_preferred_running;
5487 sgs->sum_weighted_load += weighted_cpuload(i);
5492 /* Adjust by relative CPU power of the group */
5493 sgs->group_power = group->sgp->power;
5494 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5496 if (sgs->sum_nr_running)
5497 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5499 sgs->group_weight = group->group_weight;
5501 sgs->group_imb = sg_imbalanced(group);
5502 sgs->group_capacity = sg_capacity(env, group);
5504 if (sgs->group_capacity > sgs->sum_nr_running)
5505 sgs->group_has_capacity = 1;
5509 * update_sd_pick_busiest - return 1 on busiest group
5510 * @env: The load balancing environment.
5511 * @sds: sched_domain statistics
5512 * @sg: sched_group candidate to be checked for being the busiest
5513 * @sgs: sched_group statistics
5515 * Determine if @sg is a busier group than the previously selected
5518 * Return: %true if @sg is a busier group than the previously selected
5519 * busiest group. %false otherwise.
5521 static bool update_sd_pick_busiest(struct lb_env *env,
5522 struct sd_lb_stats *sds,
5523 struct sched_group *sg,
5524 struct sg_lb_stats *sgs)
5526 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5529 if (sgs->sum_nr_running > sgs->group_capacity)
5536 * ASYM_PACKING needs to move all the work to the lowest
5537 * numbered CPUs in the group, therefore mark all groups
5538 * higher than ourself as busy.
5540 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5541 env->dst_cpu < group_first_cpu(sg)) {
5545 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5552 #ifdef CONFIG_NUMA_BALANCING
5553 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5555 if (sgs->sum_nr_running > sgs->nr_numa_running)
5557 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5562 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5564 if (rq->nr_running > rq->nr_numa_running)
5566 if (rq->nr_running > rq->nr_preferred_running)
5571 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5576 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5580 #endif /* CONFIG_NUMA_BALANCING */
5583 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5584 * @env: The load balancing environment.
5585 * @balance: Should we balance.
5586 * @sds: variable to hold the statistics for this sched_domain.
5588 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5590 struct sched_domain *child = env->sd->child;
5591 struct sched_group *sg = env->sd->groups;
5592 struct sg_lb_stats tmp_sgs;
5593 int load_idx, prefer_sibling = 0;
5595 if (child && child->flags & SD_PREFER_SIBLING)
5598 load_idx = get_sd_load_idx(env->sd, env->idle);
5601 struct sg_lb_stats *sgs = &tmp_sgs;
5604 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5607 sgs = &sds->local_stat;
5609 if (env->idle != CPU_NEWLY_IDLE ||
5610 time_after_eq(jiffies, sg->sgp->next_update))
5611 update_group_power(env->sd, env->dst_cpu);
5614 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5620 * In case the child domain prefers tasks go to siblings
5621 * first, lower the sg capacity to one so that we'll try
5622 * and move all the excess tasks away. We lower the capacity
5623 * of a group only if the local group has the capacity to fit
5624 * these excess tasks, i.e. nr_running < group_capacity. The
5625 * extra check prevents the case where you always pull from the
5626 * heaviest group when it is already under-utilized (possible
5627 * with a large weight task outweighs the tasks on the system).
5629 if (prefer_sibling && sds->local &&
5630 sds->local_stat.group_has_capacity)
5631 sgs->group_capacity = min(sgs->group_capacity, 1U);
5633 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5635 sds->busiest_stat = *sgs;
5639 /* Now, start updating sd_lb_stats */
5640 sds->total_load += sgs->group_load;
5641 sds->total_pwr += sgs->group_power;
5644 } while (sg != env->sd->groups);
5646 if (env->sd->flags & SD_NUMA)
5647 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5651 * check_asym_packing - Check to see if the group is packed into the
5654 * This is primarily intended to used at the sibling level. Some
5655 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5656 * case of POWER7, it can move to lower SMT modes only when higher
5657 * threads are idle. When in lower SMT modes, the threads will
5658 * perform better since they share less core resources. Hence when we
5659 * have idle threads, we want them to be the higher ones.
5661 * This packing function is run on idle threads. It checks to see if
5662 * the busiest CPU in this domain (core in the P7 case) has a higher
5663 * CPU number than the packing function is being run on. Here we are
5664 * assuming lower CPU number will be equivalent to lower a SMT thread
5667 * Return: 1 when packing is required and a task should be moved to
5668 * this CPU. The amount of the imbalance is returned in *imbalance.
5670 * @env: The load balancing environment.
5671 * @sds: Statistics of the sched_domain which is to be packed
5673 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5677 if (!(env->sd->flags & SD_ASYM_PACKING))
5683 busiest_cpu = group_first_cpu(sds->busiest);
5684 if (env->dst_cpu > busiest_cpu)
5687 env->imbalance = DIV_ROUND_CLOSEST(
5688 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5695 * fix_small_imbalance - Calculate the minor imbalance that exists
5696 * amongst the groups of a sched_domain, during
5698 * @env: The load balancing environment.
5699 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5702 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5704 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5705 unsigned int imbn = 2;
5706 unsigned long scaled_busy_load_per_task;
5707 struct sg_lb_stats *local, *busiest;
5709 local = &sds->local_stat;
5710 busiest = &sds->busiest_stat;
5712 if (!local->sum_nr_running)
5713 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5714 else if (busiest->load_per_task > local->load_per_task)
5717 scaled_busy_load_per_task =
5718 (busiest->load_per_task * SCHED_POWER_SCALE) /
5719 busiest->group_power;
5721 if (busiest->avg_load + scaled_busy_load_per_task >=
5722 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5723 env->imbalance = busiest->load_per_task;
5728 * OK, we don't have enough imbalance to justify moving tasks,
5729 * however we may be able to increase total CPU power used by
5733 pwr_now += busiest->group_power *
5734 min(busiest->load_per_task, busiest->avg_load);
5735 pwr_now += local->group_power *
5736 min(local->load_per_task, local->avg_load);
5737 pwr_now /= SCHED_POWER_SCALE;
5739 /* Amount of load we'd subtract */
5740 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5741 busiest->group_power;
5742 if (busiest->avg_load > tmp) {
5743 pwr_move += busiest->group_power *
5744 min(busiest->load_per_task,
5745 busiest->avg_load - tmp);
5748 /* Amount of load we'd add */
5749 if (busiest->avg_load * busiest->group_power <
5750 busiest->load_per_task * SCHED_POWER_SCALE) {
5751 tmp = (busiest->avg_load * busiest->group_power) /
5754 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5757 pwr_move += local->group_power *
5758 min(local->load_per_task, local->avg_load + tmp);
5759 pwr_move /= SCHED_POWER_SCALE;
5761 /* Move if we gain throughput */
5762 if (pwr_move > pwr_now)
5763 env->imbalance = busiest->load_per_task;
5767 * calculate_imbalance - Calculate the amount of imbalance present within the
5768 * groups of a given sched_domain during load balance.
5769 * @env: load balance environment
5770 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5772 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5774 unsigned long max_pull, load_above_capacity = ~0UL;
5775 struct sg_lb_stats *local, *busiest;
5777 local = &sds->local_stat;
5778 busiest = &sds->busiest_stat;
5780 if (busiest->group_imb) {
5782 * In the group_imb case we cannot rely on group-wide averages
5783 * to ensure cpu-load equilibrium, look at wider averages. XXX
5785 busiest->load_per_task =
5786 min(busiest->load_per_task, sds->avg_load);
5790 * In the presence of smp nice balancing, certain scenarios can have
5791 * max load less than avg load(as we skip the groups at or below
5792 * its cpu_power, while calculating max_load..)
5794 if (busiest->avg_load <= sds->avg_load ||
5795 local->avg_load >= sds->avg_load) {
5797 return fix_small_imbalance(env, sds);
5800 if (!busiest->group_imb) {
5802 * Don't want to pull so many tasks that a group would go idle.
5803 * Except of course for the group_imb case, since then we might
5804 * have to drop below capacity to reach cpu-load equilibrium.
5806 load_above_capacity =
5807 (busiest->sum_nr_running - busiest->group_capacity);
5809 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5810 load_above_capacity /= busiest->group_power;
5814 * We're trying to get all the cpus to the average_load, so we don't
5815 * want to push ourselves above the average load, nor do we wish to
5816 * reduce the max loaded cpu below the average load. At the same time,
5817 * we also don't want to reduce the group load below the group capacity
5818 * (so that we can implement power-savings policies etc). Thus we look
5819 * for the minimum possible imbalance.
5821 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5823 /* How much load to actually move to equalise the imbalance */
5824 env->imbalance = min(
5825 max_pull * busiest->group_power,
5826 (sds->avg_load - local->avg_load) * local->group_power
5827 ) / SCHED_POWER_SCALE;
5830 * if *imbalance is less than the average load per runnable task
5831 * there is no guarantee that any tasks will be moved so we'll have
5832 * a think about bumping its value to force at least one task to be
5835 if (env->imbalance < busiest->load_per_task)
5836 return fix_small_imbalance(env, sds);
5839 /******* find_busiest_group() helpers end here *********************/
5842 * find_busiest_group - Returns the busiest group within the sched_domain
5843 * if there is an imbalance. If there isn't an imbalance, and
5844 * the user has opted for power-savings, it returns a group whose
5845 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5846 * such a group exists.
5848 * Also calculates the amount of weighted load which should be moved
5849 * to restore balance.
5851 * @env: The load balancing environment.
5853 * Return: - The busiest group if imbalance exists.
5854 * - If no imbalance and user has opted for power-savings balance,
5855 * return the least loaded group whose CPUs can be
5856 * put to idle by rebalancing its tasks onto our group.
5858 static struct sched_group *find_busiest_group(struct lb_env *env)
5860 struct sg_lb_stats *local, *busiest;
5861 struct sd_lb_stats sds;
5863 init_sd_lb_stats(&sds);
5866 * Compute the various statistics relavent for load balancing at
5869 update_sd_lb_stats(env, &sds);
5870 local = &sds.local_stat;
5871 busiest = &sds.busiest_stat;
5873 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5874 check_asym_packing(env, &sds))
5877 /* There is no busy sibling group to pull tasks from */
5878 if (!sds.busiest || busiest->sum_nr_running == 0)
5881 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5884 * If the busiest group is imbalanced the below checks don't
5885 * work because they assume all things are equal, which typically
5886 * isn't true due to cpus_allowed constraints and the like.
5888 if (busiest->group_imb)
5891 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5892 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5893 !busiest->group_has_capacity)
5897 * If the local group is more busy than the selected busiest group
5898 * don't try and pull any tasks.
5900 if (local->avg_load >= busiest->avg_load)
5904 * Don't pull any tasks if this group is already above the domain
5907 if (local->avg_load >= sds.avg_load)
5910 if (env->idle == CPU_IDLE) {
5912 * This cpu is idle. If the busiest group load doesn't
5913 * have more tasks than the number of available cpu's and
5914 * there is no imbalance between this and busiest group
5915 * wrt to idle cpu's, it is balanced.
5917 if ((local->idle_cpus < busiest->idle_cpus) &&
5918 busiest->sum_nr_running <= busiest->group_weight)
5922 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5923 * imbalance_pct to be conservative.
5925 if (100 * busiest->avg_load <=
5926 env->sd->imbalance_pct * local->avg_load)
5931 /* Looks like there is an imbalance. Compute it */
5932 calculate_imbalance(env, &sds);
5941 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5943 static struct rq *find_busiest_queue(struct lb_env *env,
5944 struct sched_group *group)
5946 struct rq *busiest = NULL, *rq;
5947 unsigned long busiest_load = 0, busiest_power = 1;
5950 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5951 unsigned long power, capacity, wl;
5955 rt = fbq_classify_rq(rq);
5958 * We classify groups/runqueues into three groups:
5959 * - regular: there are !numa tasks
5960 * - remote: there are numa tasks that run on the 'wrong' node
5961 * - all: there is no distinction
5963 * In order to avoid migrating ideally placed numa tasks,
5964 * ignore those when there's better options.
5966 * If we ignore the actual busiest queue to migrate another
5967 * task, the next balance pass can still reduce the busiest
5968 * queue by moving tasks around inside the node.
5970 * If we cannot move enough load due to this classification
5971 * the next pass will adjust the group classification and
5972 * allow migration of more tasks.
5974 * Both cases only affect the total convergence complexity.
5976 if (rt > env->fbq_type)
5979 power = power_of(i);
5980 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5982 capacity = fix_small_capacity(env->sd, group);
5984 wl = weighted_cpuload(i);
5987 * When comparing with imbalance, use weighted_cpuload()
5988 * which is not scaled with the cpu power.
5990 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5994 * For the load comparisons with the other cpu's, consider
5995 * the weighted_cpuload() scaled with the cpu power, so that
5996 * the load can be moved away from the cpu that is potentially
5997 * running at a lower capacity.
5999 * Thus we're looking for max(wl_i / power_i), crosswise
6000 * multiplication to rid ourselves of the division works out
6001 * to: wl_i * power_j > wl_j * power_i; where j is our
6004 if (wl * busiest_power > busiest_load * power) {
6006 busiest_power = power;
6015 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6016 * so long as it is large enough.
6018 #define MAX_PINNED_INTERVAL 512
6020 /* Working cpumask for load_balance and load_balance_newidle. */
6021 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6023 static int need_active_balance(struct lb_env *env)
6025 struct sched_domain *sd = env->sd;
6027 if (env->idle == CPU_NEWLY_IDLE) {
6030 * ASYM_PACKING needs to force migrate tasks from busy but
6031 * higher numbered CPUs in order to pack all tasks in the
6032 * lowest numbered CPUs.
6034 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6038 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6041 static int active_load_balance_cpu_stop(void *data);
6043 static int should_we_balance(struct lb_env *env)
6045 struct sched_group *sg = env->sd->groups;
6046 struct cpumask *sg_cpus, *sg_mask;
6047 int cpu, balance_cpu = -1;
6050 * In the newly idle case, we will allow all the cpu's
6051 * to do the newly idle load balance.
6053 if (env->idle == CPU_NEWLY_IDLE)
6056 sg_cpus = sched_group_cpus(sg);
6057 sg_mask = sched_group_mask(sg);
6058 /* Try to find first idle cpu */
6059 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6060 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6067 if (balance_cpu == -1)
6068 balance_cpu = group_balance_cpu(sg);
6071 * First idle cpu or the first cpu(busiest) in this sched group
6072 * is eligible for doing load balancing at this and above domains.
6074 return balance_cpu == env->dst_cpu;
6078 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6079 * tasks if there is an imbalance.
6081 static int load_balance(int this_cpu, struct rq *this_rq,
6082 struct sched_domain *sd, enum cpu_idle_type idle,
6083 int *continue_balancing)
6085 int ld_moved, cur_ld_moved, active_balance = 0;
6086 struct sched_domain *sd_parent = sd->parent;
6087 struct sched_group *group;
6089 unsigned long flags;
6090 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6092 struct lb_env env = {
6094 .dst_cpu = this_cpu,
6096 .dst_grpmask = sched_group_cpus(sd->groups),
6098 .loop_break = sched_nr_migrate_break,
6104 * For NEWLY_IDLE load_balancing, we don't need to consider
6105 * other cpus in our group
6107 if (idle == CPU_NEWLY_IDLE)
6108 env.dst_grpmask = NULL;
6110 cpumask_copy(cpus, cpu_active_mask);
6112 schedstat_inc(sd, lb_count[idle]);
6115 if (!should_we_balance(&env)) {
6116 *continue_balancing = 0;
6120 group = find_busiest_group(&env);
6122 schedstat_inc(sd, lb_nobusyg[idle]);
6126 busiest = find_busiest_queue(&env, group);
6128 schedstat_inc(sd, lb_nobusyq[idle]);
6132 BUG_ON(busiest == env.dst_rq);
6134 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6137 if (busiest->nr_running > 1) {
6139 * Attempt to move tasks. If find_busiest_group has found
6140 * an imbalance but busiest->nr_running <= 1, the group is
6141 * still unbalanced. ld_moved simply stays zero, so it is
6142 * correctly treated as an imbalance.
6144 env.flags |= LBF_ALL_PINNED;
6145 env.src_cpu = busiest->cpu;
6146 env.src_rq = busiest;
6147 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6150 local_irq_save(flags);
6151 double_rq_lock(env.dst_rq, busiest);
6154 * cur_ld_moved - load moved in current iteration
6155 * ld_moved - cumulative load moved across iterations
6157 cur_ld_moved = move_tasks(&env);
6158 ld_moved += cur_ld_moved;
6159 double_rq_unlock(env.dst_rq, busiest);
6160 local_irq_restore(flags);
6163 * some other cpu did the load balance for us.
6165 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6166 resched_cpu(env.dst_cpu);
6168 if (env.flags & LBF_NEED_BREAK) {
6169 env.flags &= ~LBF_NEED_BREAK;
6174 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6175 * us and move them to an alternate dst_cpu in our sched_group
6176 * where they can run. The upper limit on how many times we
6177 * iterate on same src_cpu is dependent on number of cpus in our
6180 * This changes load balance semantics a bit on who can move
6181 * load to a given_cpu. In addition to the given_cpu itself
6182 * (or a ilb_cpu acting on its behalf where given_cpu is
6183 * nohz-idle), we now have balance_cpu in a position to move
6184 * load to given_cpu. In rare situations, this may cause
6185 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6186 * _independently_ and at _same_ time to move some load to
6187 * given_cpu) causing exceess load to be moved to given_cpu.
6188 * This however should not happen so much in practice and
6189 * moreover subsequent load balance cycles should correct the
6190 * excess load moved.
6192 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6194 /* Prevent to re-select dst_cpu via env's cpus */
6195 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6197 env.dst_rq = cpu_rq(env.new_dst_cpu);
6198 env.dst_cpu = env.new_dst_cpu;
6199 env.flags &= ~LBF_DST_PINNED;
6201 env.loop_break = sched_nr_migrate_break;
6204 * Go back to "more_balance" rather than "redo" since we
6205 * need to continue with same src_cpu.
6211 * We failed to reach balance because of affinity.
6214 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6216 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6217 *group_imbalance = 1;
6218 } else if (*group_imbalance)
6219 *group_imbalance = 0;
6222 /* All tasks on this runqueue were pinned by CPU affinity */
6223 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6224 cpumask_clear_cpu(cpu_of(busiest), cpus);
6225 if (!cpumask_empty(cpus)) {
6227 env.loop_break = sched_nr_migrate_break;
6235 schedstat_inc(sd, lb_failed[idle]);
6237 * Increment the failure counter only on periodic balance.
6238 * We do not want newidle balance, which can be very
6239 * frequent, pollute the failure counter causing
6240 * excessive cache_hot migrations and active balances.
6242 if (idle != CPU_NEWLY_IDLE)
6243 sd->nr_balance_failed++;
6245 if (need_active_balance(&env)) {
6246 raw_spin_lock_irqsave(&busiest->lock, flags);
6248 /* don't kick the active_load_balance_cpu_stop,
6249 * if the curr task on busiest cpu can't be
6252 if (!cpumask_test_cpu(this_cpu,
6253 tsk_cpus_allowed(busiest->curr))) {
6254 raw_spin_unlock_irqrestore(&busiest->lock,
6256 env.flags |= LBF_ALL_PINNED;
6257 goto out_one_pinned;
6261 * ->active_balance synchronizes accesses to
6262 * ->active_balance_work. Once set, it's cleared
6263 * only after active load balance is finished.
6265 if (!busiest->active_balance) {
6266 busiest->active_balance = 1;
6267 busiest->push_cpu = this_cpu;
6270 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6272 if (active_balance) {
6273 stop_one_cpu_nowait(cpu_of(busiest),
6274 active_load_balance_cpu_stop, busiest,
6275 &busiest->active_balance_work);
6279 * We've kicked active balancing, reset the failure
6282 sd->nr_balance_failed = sd->cache_nice_tries+1;
6285 sd->nr_balance_failed = 0;
6287 if (likely(!active_balance)) {
6288 /* We were unbalanced, so reset the balancing interval */
6289 sd->balance_interval = sd->min_interval;
6292 * If we've begun active balancing, start to back off. This
6293 * case may not be covered by the all_pinned logic if there
6294 * is only 1 task on the busy runqueue (because we don't call
6297 if (sd->balance_interval < sd->max_interval)
6298 sd->balance_interval *= 2;
6304 schedstat_inc(sd, lb_balanced[idle]);
6306 sd->nr_balance_failed = 0;
6309 /* tune up the balancing interval */
6310 if (((env.flags & LBF_ALL_PINNED) &&
6311 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6312 (sd->balance_interval < sd->max_interval))
6313 sd->balance_interval *= 2;
6321 * idle_balance is called by schedule() if this_cpu is about to become
6322 * idle. Attempts to pull tasks from other CPUs.
6324 void idle_balance(int this_cpu, struct rq *this_rq)
6326 struct sched_domain *sd;
6327 int pulled_task = 0;
6328 unsigned long next_balance = jiffies + HZ;
6331 this_rq->idle_stamp = rq_clock(this_rq);
6333 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6337 * Drop the rq->lock, but keep IRQ/preempt disabled.
6339 raw_spin_unlock(&this_rq->lock);
6341 update_blocked_averages(this_cpu);
6343 for_each_domain(this_cpu, sd) {
6344 unsigned long interval;
6345 int continue_balancing = 1;
6346 u64 t0, domain_cost;
6348 if (!(sd->flags & SD_LOAD_BALANCE))
6351 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6354 if (sd->flags & SD_BALANCE_NEWIDLE) {
6355 t0 = sched_clock_cpu(this_cpu);
6357 /* If we've pulled tasks over stop searching: */
6358 pulled_task = load_balance(this_cpu, this_rq,
6360 &continue_balancing);
6362 domain_cost = sched_clock_cpu(this_cpu) - t0;
6363 if (domain_cost > sd->max_newidle_lb_cost)
6364 sd->max_newidle_lb_cost = domain_cost;
6366 curr_cost += domain_cost;
6369 interval = msecs_to_jiffies(sd->balance_interval);
6370 if (time_after(next_balance, sd->last_balance + interval))
6371 next_balance = sd->last_balance + interval;
6373 this_rq->idle_stamp = 0;
6379 raw_spin_lock(&this_rq->lock);
6381 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6383 * We are going idle. next_balance may be set based on
6384 * a busy processor. So reset next_balance.
6386 this_rq->next_balance = next_balance;
6389 if (curr_cost > this_rq->max_idle_balance_cost)
6390 this_rq->max_idle_balance_cost = curr_cost;
6394 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6395 * running tasks off the busiest CPU onto idle CPUs. It requires at
6396 * least 1 task to be running on each physical CPU where possible, and
6397 * avoids physical / logical imbalances.
6399 static int active_load_balance_cpu_stop(void *data)
6401 struct rq *busiest_rq = data;
6402 int busiest_cpu = cpu_of(busiest_rq);
6403 int target_cpu = busiest_rq->push_cpu;
6404 struct rq *target_rq = cpu_rq(target_cpu);
6405 struct sched_domain *sd;
6407 raw_spin_lock_irq(&busiest_rq->lock);
6409 /* make sure the requested cpu hasn't gone down in the meantime */
6410 if (unlikely(busiest_cpu != smp_processor_id() ||
6411 !busiest_rq->active_balance))
6414 /* Is there any task to move? */
6415 if (busiest_rq->nr_running <= 1)
6419 * This condition is "impossible", if it occurs
6420 * we need to fix it. Originally reported by
6421 * Bjorn Helgaas on a 128-cpu setup.
6423 BUG_ON(busiest_rq == target_rq);
6425 /* move a task from busiest_rq to target_rq */
6426 double_lock_balance(busiest_rq, target_rq);
6428 /* Search for an sd spanning us and the target CPU. */
6430 for_each_domain(target_cpu, sd) {
6431 if ((sd->flags & SD_LOAD_BALANCE) &&
6432 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6437 struct lb_env env = {
6439 .dst_cpu = target_cpu,
6440 .dst_rq = target_rq,
6441 .src_cpu = busiest_rq->cpu,
6442 .src_rq = busiest_rq,
6446 schedstat_inc(sd, alb_count);
6448 if (move_one_task(&env))
6449 schedstat_inc(sd, alb_pushed);
6451 schedstat_inc(sd, alb_failed);
6454 double_unlock_balance(busiest_rq, target_rq);
6456 busiest_rq->active_balance = 0;
6457 raw_spin_unlock_irq(&busiest_rq->lock);
6461 #ifdef CONFIG_NO_HZ_COMMON
6463 * idle load balancing details
6464 * - When one of the busy CPUs notice that there may be an idle rebalancing
6465 * needed, they will kick the idle load balancer, which then does idle
6466 * load balancing for all the idle CPUs.
6469 cpumask_var_t idle_cpus_mask;
6471 unsigned long next_balance; /* in jiffy units */
6472 } nohz ____cacheline_aligned;
6474 static inline int find_new_ilb(int call_cpu)
6476 int ilb = cpumask_first(nohz.idle_cpus_mask);
6478 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6485 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6486 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6487 * CPU (if there is one).
6489 static void nohz_balancer_kick(int cpu)
6493 nohz.next_balance++;
6495 ilb_cpu = find_new_ilb(cpu);
6497 if (ilb_cpu >= nr_cpu_ids)
6500 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6503 * Use smp_send_reschedule() instead of resched_cpu().
6504 * This way we generate a sched IPI on the target cpu which
6505 * is idle. And the softirq performing nohz idle load balance
6506 * will be run before returning from the IPI.
6508 smp_send_reschedule(ilb_cpu);
6512 static inline void nohz_balance_exit_idle(int cpu)
6514 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6515 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6516 atomic_dec(&nohz.nr_cpus);
6517 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6521 static inline void set_cpu_sd_state_busy(void)
6523 struct sched_domain *sd;
6526 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6528 if (!sd || !sd->nohz_idle)
6532 for (; sd; sd = sd->parent)
6533 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6538 void set_cpu_sd_state_idle(void)
6540 struct sched_domain *sd;
6543 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6545 if (!sd || sd->nohz_idle)
6549 for (; sd; sd = sd->parent)
6550 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6556 * This routine will record that the cpu is going idle with tick stopped.
6557 * This info will be used in performing idle load balancing in the future.
6559 void nohz_balance_enter_idle(int cpu)
6562 * If this cpu is going down, then nothing needs to be done.
6564 if (!cpu_active(cpu))
6567 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6570 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6571 atomic_inc(&nohz.nr_cpus);
6572 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6575 static int sched_ilb_notifier(struct notifier_block *nfb,
6576 unsigned long action, void *hcpu)
6578 switch (action & ~CPU_TASKS_FROZEN) {
6580 nohz_balance_exit_idle(smp_processor_id());
6588 static DEFINE_SPINLOCK(balancing);
6591 * Scale the max load_balance interval with the number of CPUs in the system.
6592 * This trades load-balance latency on larger machines for less cross talk.
6594 void update_max_interval(void)
6596 max_load_balance_interval = HZ*num_online_cpus()/10;
6600 * It checks each scheduling domain to see if it is due to be balanced,
6601 * and initiates a balancing operation if so.
6603 * Balancing parameters are set up in init_sched_domains.
6605 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6607 int continue_balancing = 1;
6608 struct rq *rq = cpu_rq(cpu);
6609 unsigned long interval;
6610 struct sched_domain *sd;
6611 /* Earliest time when we have to do rebalance again */
6612 unsigned long next_balance = jiffies + 60*HZ;
6613 int update_next_balance = 0;
6614 int need_serialize, need_decay = 0;
6617 update_blocked_averages(cpu);
6620 for_each_domain(cpu, sd) {
6622 * Decay the newidle max times here because this is a regular
6623 * visit to all the domains. Decay ~1% per second.
6625 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6626 sd->max_newidle_lb_cost =
6627 (sd->max_newidle_lb_cost * 253) / 256;
6628 sd->next_decay_max_lb_cost = jiffies + HZ;
6631 max_cost += sd->max_newidle_lb_cost;
6633 if (!(sd->flags & SD_LOAD_BALANCE))
6637 * Stop the load balance at this level. There is another
6638 * CPU in our sched group which is doing load balancing more
6641 if (!continue_balancing) {
6647 interval = sd->balance_interval;
6648 if (idle != CPU_IDLE)
6649 interval *= sd->busy_factor;
6651 /* scale ms to jiffies */
6652 interval = msecs_to_jiffies(interval);
6653 interval = clamp(interval, 1UL, max_load_balance_interval);
6655 need_serialize = sd->flags & SD_SERIALIZE;
6657 if (need_serialize) {
6658 if (!spin_trylock(&balancing))
6662 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6663 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6665 * The LBF_DST_PINNED logic could have changed
6666 * env->dst_cpu, so we can't know our idle
6667 * state even if we migrated tasks. Update it.
6669 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6671 sd->last_balance = jiffies;
6674 spin_unlock(&balancing);
6676 if (time_after(next_balance, sd->last_balance + interval)) {
6677 next_balance = sd->last_balance + interval;
6678 update_next_balance = 1;
6683 * Ensure the rq-wide value also decays but keep it at a
6684 * reasonable floor to avoid funnies with rq->avg_idle.
6686 rq->max_idle_balance_cost =
6687 max((u64)sysctl_sched_migration_cost, max_cost);
6692 * next_balance will be updated only when there is a need.
6693 * When the cpu is attached to null domain for ex, it will not be
6696 if (likely(update_next_balance))
6697 rq->next_balance = next_balance;
6700 #ifdef CONFIG_NO_HZ_COMMON
6702 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6703 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6705 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6707 struct rq *this_rq = cpu_rq(this_cpu);
6711 if (idle != CPU_IDLE ||
6712 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6715 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6716 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6720 * If this cpu gets work to do, stop the load balancing
6721 * work being done for other cpus. Next load
6722 * balancing owner will pick it up.
6727 rq = cpu_rq(balance_cpu);
6729 raw_spin_lock_irq(&rq->lock);
6730 update_rq_clock(rq);
6731 update_idle_cpu_load(rq);
6732 raw_spin_unlock_irq(&rq->lock);
6734 rebalance_domains(balance_cpu, CPU_IDLE);
6736 if (time_after(this_rq->next_balance, rq->next_balance))
6737 this_rq->next_balance = rq->next_balance;
6739 nohz.next_balance = this_rq->next_balance;
6741 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6745 * Current heuristic for kicking the idle load balancer in the presence
6746 * of an idle cpu is the system.
6747 * - This rq has more than one task.
6748 * - At any scheduler domain level, this cpu's scheduler group has multiple
6749 * busy cpu's exceeding the group's power.
6750 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6751 * domain span are idle.
6753 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6755 unsigned long now = jiffies;
6756 struct sched_domain *sd;
6758 if (unlikely(idle_cpu(cpu)))
6762 * We may be recently in ticked or tickless idle mode. At the first
6763 * busy tick after returning from idle, we will update the busy stats.
6765 set_cpu_sd_state_busy();
6766 nohz_balance_exit_idle(cpu);
6769 * None are in tickless mode and hence no need for NOHZ idle load
6772 if (likely(!atomic_read(&nohz.nr_cpus)))
6775 if (time_before(now, nohz.next_balance))
6778 if (rq->nr_running >= 2)
6782 for_each_domain(cpu, sd) {
6783 struct sched_group *sg = sd->groups;
6784 struct sched_group_power *sgp = sg->sgp;
6785 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6787 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6788 goto need_kick_unlock;
6790 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6791 && (cpumask_first_and(nohz.idle_cpus_mask,
6792 sched_domain_span(sd)) < cpu))
6793 goto need_kick_unlock;
6795 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6807 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6811 * run_rebalance_domains is triggered when needed from the scheduler tick.
6812 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6814 static void run_rebalance_domains(struct softirq_action *h)
6816 int this_cpu = smp_processor_id();
6817 struct rq *this_rq = cpu_rq(this_cpu);
6818 enum cpu_idle_type idle = this_rq->idle_balance ?
6819 CPU_IDLE : CPU_NOT_IDLE;
6821 rebalance_domains(this_cpu, idle);
6824 * If this cpu has a pending nohz_balance_kick, then do the
6825 * balancing on behalf of the other idle cpus whose ticks are
6828 nohz_idle_balance(this_cpu, idle);
6831 static inline int on_null_domain(int cpu)
6833 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6837 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6839 void trigger_load_balance(struct rq *rq, int cpu)
6841 /* Don't need to rebalance while attached to NULL domain */
6842 if (time_after_eq(jiffies, rq->next_balance) &&
6843 likely(!on_null_domain(cpu)))
6844 raise_softirq(SCHED_SOFTIRQ);
6845 #ifdef CONFIG_NO_HZ_COMMON
6846 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6847 nohz_balancer_kick(cpu);
6851 static void rq_online_fair(struct rq *rq)
6856 static void rq_offline_fair(struct rq *rq)
6860 /* Ensure any throttled groups are reachable by pick_next_task */
6861 unthrottle_offline_cfs_rqs(rq);
6864 #endif /* CONFIG_SMP */
6867 * scheduler tick hitting a task of our scheduling class:
6869 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6871 struct cfs_rq *cfs_rq;
6872 struct sched_entity *se = &curr->se;
6874 for_each_sched_entity(se) {
6875 cfs_rq = cfs_rq_of(se);
6876 entity_tick(cfs_rq, se, queued);
6879 if (numabalancing_enabled)
6880 task_tick_numa(rq, curr);
6882 update_rq_runnable_avg(rq, 1);
6886 * called on fork with the child task as argument from the parent's context
6887 * - child not yet on the tasklist
6888 * - preemption disabled
6890 static void task_fork_fair(struct task_struct *p)
6892 struct cfs_rq *cfs_rq;
6893 struct sched_entity *se = &p->se, *curr;
6894 int this_cpu = smp_processor_id();
6895 struct rq *rq = this_rq();
6896 unsigned long flags;
6898 raw_spin_lock_irqsave(&rq->lock, flags);
6900 update_rq_clock(rq);
6902 cfs_rq = task_cfs_rq(current);
6903 curr = cfs_rq->curr;
6906 * Not only the cpu but also the task_group of the parent might have
6907 * been changed after parent->se.parent,cfs_rq were copied to
6908 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6909 * of child point to valid ones.
6912 __set_task_cpu(p, this_cpu);
6915 update_curr(cfs_rq);
6918 se->vruntime = curr->vruntime;
6919 place_entity(cfs_rq, se, 1);
6921 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6923 * Upon rescheduling, sched_class::put_prev_task() will place
6924 * 'current' within the tree based on its new key value.
6926 swap(curr->vruntime, se->vruntime);
6927 resched_task(rq->curr);
6930 se->vruntime -= cfs_rq->min_vruntime;
6932 raw_spin_unlock_irqrestore(&rq->lock, flags);
6936 * Priority of the task has changed. Check to see if we preempt
6940 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6946 * Reschedule if we are currently running on this runqueue and
6947 * our priority decreased, or if we are not currently running on
6948 * this runqueue and our priority is higher than the current's
6950 if (rq->curr == p) {
6951 if (p->prio > oldprio)
6952 resched_task(rq->curr);
6954 check_preempt_curr(rq, p, 0);
6957 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6959 struct sched_entity *se = &p->se;
6960 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6963 * Ensure the task's vruntime is normalized, so that when its
6964 * switched back to the fair class the enqueue_entity(.flags=0) will
6965 * do the right thing.
6967 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6968 * have normalized the vruntime, if it was !on_rq, then only when
6969 * the task is sleeping will it still have non-normalized vruntime.
6971 if (!se->on_rq && p->state != TASK_RUNNING) {
6973 * Fix up our vruntime so that the current sleep doesn't
6974 * cause 'unlimited' sleep bonus.
6976 place_entity(cfs_rq, se, 0);
6977 se->vruntime -= cfs_rq->min_vruntime;
6982 * Remove our load from contribution when we leave sched_fair
6983 * and ensure we don't carry in an old decay_count if we
6986 if (se->avg.decay_count) {
6987 __synchronize_entity_decay(se);
6988 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6994 * We switched to the sched_fair class.
6996 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7002 * We were most likely switched from sched_rt, so
7003 * kick off the schedule if running, otherwise just see
7004 * if we can still preempt the current task.
7007 resched_task(rq->curr);
7009 check_preempt_curr(rq, p, 0);
7012 /* Account for a task changing its policy or group.
7014 * This routine is mostly called to set cfs_rq->curr field when a task
7015 * migrates between groups/classes.
7017 static void set_curr_task_fair(struct rq *rq)
7019 struct sched_entity *se = &rq->curr->se;
7021 for_each_sched_entity(se) {
7022 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7024 set_next_entity(cfs_rq, se);
7025 /* ensure bandwidth has been allocated on our new cfs_rq */
7026 account_cfs_rq_runtime(cfs_rq, 0);
7030 void init_cfs_rq(struct cfs_rq *cfs_rq)
7032 cfs_rq->tasks_timeline = RB_ROOT;
7033 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7034 #ifndef CONFIG_64BIT
7035 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7038 atomic64_set(&cfs_rq->decay_counter, 1);
7039 atomic_long_set(&cfs_rq->removed_load, 0);
7043 #ifdef CONFIG_FAIR_GROUP_SCHED
7044 static void task_move_group_fair(struct task_struct *p, int on_rq)
7046 struct cfs_rq *cfs_rq;
7048 * If the task was not on the rq at the time of this cgroup movement
7049 * it must have been asleep, sleeping tasks keep their ->vruntime
7050 * absolute on their old rq until wakeup (needed for the fair sleeper
7051 * bonus in place_entity()).
7053 * If it was on the rq, we've just 'preempted' it, which does convert
7054 * ->vruntime to a relative base.
7056 * Make sure both cases convert their relative position when migrating
7057 * to another cgroup's rq. This does somewhat interfere with the
7058 * fair sleeper stuff for the first placement, but who cares.
7061 * When !on_rq, vruntime of the task has usually NOT been normalized.
7062 * But there are some cases where it has already been normalized:
7064 * - Moving a forked child which is waiting for being woken up by
7065 * wake_up_new_task().
7066 * - Moving a task which has been woken up by try_to_wake_up() and
7067 * waiting for actually being woken up by sched_ttwu_pending().
7069 * To prevent boost or penalty in the new cfs_rq caused by delta
7070 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7072 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7076 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7077 set_task_rq(p, task_cpu(p));
7079 cfs_rq = cfs_rq_of(&p->se);
7080 p->se.vruntime += cfs_rq->min_vruntime;
7083 * migrate_task_rq_fair() will have removed our previous
7084 * contribution, but we must synchronize for ongoing future
7087 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7088 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7093 void free_fair_sched_group(struct task_group *tg)
7097 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7099 for_each_possible_cpu(i) {
7101 kfree(tg->cfs_rq[i]);
7110 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7112 struct cfs_rq *cfs_rq;
7113 struct sched_entity *se;
7116 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7119 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7123 tg->shares = NICE_0_LOAD;
7125 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7127 for_each_possible_cpu(i) {
7128 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7129 GFP_KERNEL, cpu_to_node(i));
7133 se = kzalloc_node(sizeof(struct sched_entity),
7134 GFP_KERNEL, cpu_to_node(i));
7138 init_cfs_rq(cfs_rq);
7139 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7150 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7152 struct rq *rq = cpu_rq(cpu);
7153 unsigned long flags;
7156 * Only empty task groups can be destroyed; so we can speculatively
7157 * check on_list without danger of it being re-added.
7159 if (!tg->cfs_rq[cpu]->on_list)
7162 raw_spin_lock_irqsave(&rq->lock, flags);
7163 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7164 raw_spin_unlock_irqrestore(&rq->lock, flags);
7167 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7168 struct sched_entity *se, int cpu,
7169 struct sched_entity *parent)
7171 struct rq *rq = cpu_rq(cpu);
7175 init_cfs_rq_runtime(cfs_rq);
7177 tg->cfs_rq[cpu] = cfs_rq;
7180 /* se could be NULL for root_task_group */
7185 se->cfs_rq = &rq->cfs;
7187 se->cfs_rq = parent->my_q;
7190 update_load_set(&se->load, 0);
7191 se->parent = parent;
7194 static DEFINE_MUTEX(shares_mutex);
7196 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7199 unsigned long flags;
7202 * We can't change the weight of the root cgroup.
7207 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7209 mutex_lock(&shares_mutex);
7210 if (tg->shares == shares)
7213 tg->shares = shares;
7214 for_each_possible_cpu(i) {
7215 struct rq *rq = cpu_rq(i);
7216 struct sched_entity *se;
7219 /* Propagate contribution to hierarchy */
7220 raw_spin_lock_irqsave(&rq->lock, flags);
7222 /* Possible calls to update_curr() need rq clock */
7223 update_rq_clock(rq);
7224 for_each_sched_entity(se)
7225 update_cfs_shares(group_cfs_rq(se));
7226 raw_spin_unlock_irqrestore(&rq->lock, flags);
7230 mutex_unlock(&shares_mutex);
7233 #else /* CONFIG_FAIR_GROUP_SCHED */
7235 void free_fair_sched_group(struct task_group *tg) { }
7237 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7242 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7244 #endif /* CONFIG_FAIR_GROUP_SCHED */
7247 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7249 struct sched_entity *se = &task->se;
7250 unsigned int rr_interval = 0;
7253 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7256 if (rq->cfs.load.weight)
7257 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7263 * All the scheduling class methods:
7265 const struct sched_class fair_sched_class = {
7266 .next = &idle_sched_class,
7267 .enqueue_task = enqueue_task_fair,
7268 .dequeue_task = dequeue_task_fair,
7269 .yield_task = yield_task_fair,
7270 .yield_to_task = yield_to_task_fair,
7272 .check_preempt_curr = check_preempt_wakeup,
7274 .pick_next_task = pick_next_task_fair,
7275 .put_prev_task = put_prev_task_fair,
7278 .select_task_rq = select_task_rq_fair,
7279 .migrate_task_rq = migrate_task_rq_fair,
7281 .rq_online = rq_online_fair,
7282 .rq_offline = rq_offline_fair,
7284 .task_waking = task_waking_fair,
7287 .set_curr_task = set_curr_task_fair,
7288 .task_tick = task_tick_fair,
7289 .task_fork = task_fork_fair,
7291 .prio_changed = prio_changed_fair,
7292 .switched_from = switched_from_fair,
7293 .switched_to = switched_to_fair,
7295 .get_rr_interval = get_rr_interval_fair,
7297 #ifdef CONFIG_FAIR_GROUP_SCHED
7298 .task_move_group = task_move_group_fair,
7302 #ifdef CONFIG_SCHED_DEBUG
7303 void print_cfs_stats(struct seq_file *m, int cpu)
7305 struct cfs_rq *cfs_rq;
7308 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7309 print_cfs_rq(m, cpu, cfs_rq);
7314 __init void init_sched_fair_class(void)
7317 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7319 #ifdef CONFIG_NO_HZ_COMMON
7320 nohz.next_balance = jiffies;
7321 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7322 cpu_notifier(sched_ilb_notifier, 0);