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))
1538 double_lock(&my_grp->lock, &grp->lock);
1540 for (i = 0; i < 2*nr_node_ids; i++) {
1541 my_grp->faults[i] -= p->numa_faults[i];
1542 grp->faults[i] += p->numa_faults[i];
1544 my_grp->total_faults -= p->total_numa_faults;
1545 grp->total_faults += p->total_numa_faults;
1547 list_move(&p->numa_entry, &grp->task_list);
1551 spin_unlock(&my_grp->lock);
1552 spin_unlock(&grp->lock);
1554 rcu_assign_pointer(p->numa_group, grp);
1556 put_numa_group(my_grp);
1564 void task_numa_free(struct task_struct *p)
1566 struct numa_group *grp = p->numa_group;
1568 void *numa_faults = p->numa_faults;
1571 spin_lock(&grp->lock);
1572 for (i = 0; i < 2*nr_node_ids; i++)
1573 grp->faults[i] -= p->numa_faults[i];
1574 grp->total_faults -= p->total_numa_faults;
1576 list_del(&p->numa_entry);
1578 spin_unlock(&grp->lock);
1579 rcu_assign_pointer(p->numa_group, NULL);
1580 put_numa_group(grp);
1583 p->numa_faults = NULL;
1584 p->numa_faults_buffer = NULL;
1589 * Got a PROT_NONE fault for a page on @node.
1591 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1593 struct task_struct *p = current;
1594 bool migrated = flags & TNF_MIGRATED;
1597 if (!numabalancing_enabled)
1600 /* for example, ksmd faulting in a user's mm */
1604 /* Do not worry about placement if exiting */
1605 if (p->state == TASK_DEAD)
1608 /* Allocate buffer to track faults on a per-node basis */
1609 if (unlikely(!p->numa_faults)) {
1610 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1612 /* numa_faults and numa_faults_buffer share the allocation */
1613 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1614 if (!p->numa_faults)
1617 BUG_ON(p->numa_faults_buffer);
1618 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1619 p->total_numa_faults = 0;
1620 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1624 * First accesses are treated as private, otherwise consider accesses
1625 * to be private if the accessing pid has not changed
1627 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1630 priv = cpupid_match_pid(p, last_cpupid);
1631 if (!priv && !(flags & TNF_NO_GROUP))
1632 task_numa_group(p, last_cpupid, flags, &priv);
1635 task_numa_placement(p);
1638 * Retry task to preferred node migration periodically, in case it
1639 * case it previously failed, or the scheduler moved us.
1641 if (time_after(jiffies, p->numa_migrate_retry))
1642 numa_migrate_preferred(p);
1645 p->numa_pages_migrated += pages;
1647 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1648 p->numa_faults_locality[!!(flags & TNF_FAULT_LOCAL)] += pages;
1651 static void reset_ptenuma_scan(struct task_struct *p)
1653 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1654 p->mm->numa_scan_offset = 0;
1658 * The expensive part of numa migration is done from task_work context.
1659 * Triggered from task_tick_numa().
1661 void task_numa_work(struct callback_head *work)
1663 unsigned long migrate, next_scan, now = jiffies;
1664 struct task_struct *p = current;
1665 struct mm_struct *mm = p->mm;
1666 struct vm_area_struct *vma;
1667 unsigned long start, end;
1668 unsigned long nr_pte_updates = 0;
1671 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1673 work->next = work; /* protect against double add */
1675 * Who cares about NUMA placement when they're dying.
1677 * NOTE: make sure not to dereference p->mm before this check,
1678 * exit_task_work() happens _after_ exit_mm() so we could be called
1679 * without p->mm even though we still had it when we enqueued this
1682 if (p->flags & PF_EXITING)
1685 if (!mm->numa_next_scan) {
1686 mm->numa_next_scan = now +
1687 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1691 * Enforce maximal scan/migration frequency..
1693 migrate = mm->numa_next_scan;
1694 if (time_before(now, migrate))
1697 if (p->numa_scan_period == 0) {
1698 p->numa_scan_period_max = task_scan_max(p);
1699 p->numa_scan_period = task_scan_min(p);
1702 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1703 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1707 * Delay this task enough that another task of this mm will likely win
1708 * the next time around.
1710 p->node_stamp += 2 * TICK_NSEC;
1712 start = mm->numa_scan_offset;
1713 pages = sysctl_numa_balancing_scan_size;
1714 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1718 down_read(&mm->mmap_sem);
1719 vma = find_vma(mm, start);
1721 reset_ptenuma_scan(p);
1725 for (; vma; vma = vma->vm_next) {
1726 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1730 * Shared library pages mapped by multiple processes are not
1731 * migrated as it is expected they are cache replicated. Avoid
1732 * hinting faults in read-only file-backed mappings or the vdso
1733 * as migrating the pages will be of marginal benefit.
1736 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1740 start = max(start, vma->vm_start);
1741 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1742 end = min(end, vma->vm_end);
1743 nr_pte_updates += change_prot_numa(vma, start, end);
1746 * Scan sysctl_numa_balancing_scan_size but ensure that
1747 * at least one PTE is updated so that unused virtual
1748 * address space is quickly skipped.
1751 pages -= (end - start) >> PAGE_SHIFT;
1756 } while (end != vma->vm_end);
1761 * It is possible to reach the end of the VMA list but the last few
1762 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1763 * would find the !migratable VMA on the next scan but not reset the
1764 * scanner to the start so check it now.
1767 mm->numa_scan_offset = start;
1769 reset_ptenuma_scan(p);
1770 up_read(&mm->mmap_sem);
1774 * Drive the periodic memory faults..
1776 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1778 struct callback_head *work = &curr->numa_work;
1782 * We don't care about NUMA placement if we don't have memory.
1784 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1788 * Using runtime rather than walltime has the dual advantage that
1789 * we (mostly) drive the selection from busy threads and that the
1790 * task needs to have done some actual work before we bother with
1793 now = curr->se.sum_exec_runtime;
1794 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1796 if (now - curr->node_stamp > period) {
1797 if (!curr->node_stamp)
1798 curr->numa_scan_period = task_scan_min(curr);
1799 curr->node_stamp += period;
1801 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1802 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1803 task_work_add(curr, work, true);
1808 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1812 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1816 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1819 #endif /* CONFIG_NUMA_BALANCING */
1822 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1824 update_load_add(&cfs_rq->load, se->load.weight);
1825 if (!parent_entity(se))
1826 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1828 if (entity_is_task(se)) {
1829 struct rq *rq = rq_of(cfs_rq);
1831 account_numa_enqueue(rq, task_of(se));
1832 list_add(&se->group_node, &rq->cfs_tasks);
1835 cfs_rq->nr_running++;
1839 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1841 update_load_sub(&cfs_rq->load, se->load.weight);
1842 if (!parent_entity(se))
1843 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1844 if (entity_is_task(se)) {
1845 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
1846 list_del_init(&se->group_node);
1848 cfs_rq->nr_running--;
1851 #ifdef CONFIG_FAIR_GROUP_SCHED
1853 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1858 * Use this CPU's actual weight instead of the last load_contribution
1859 * to gain a more accurate current total weight. See
1860 * update_cfs_rq_load_contribution().
1862 tg_weight = atomic_long_read(&tg->load_avg);
1863 tg_weight -= cfs_rq->tg_load_contrib;
1864 tg_weight += cfs_rq->load.weight;
1869 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1871 long tg_weight, load, shares;
1873 tg_weight = calc_tg_weight(tg, cfs_rq);
1874 load = cfs_rq->load.weight;
1876 shares = (tg->shares * load);
1878 shares /= tg_weight;
1880 if (shares < MIN_SHARES)
1881 shares = MIN_SHARES;
1882 if (shares > tg->shares)
1883 shares = tg->shares;
1887 # else /* CONFIG_SMP */
1888 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1892 # endif /* CONFIG_SMP */
1893 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1894 unsigned long weight)
1897 /* commit outstanding execution time */
1898 if (cfs_rq->curr == se)
1899 update_curr(cfs_rq);
1900 account_entity_dequeue(cfs_rq, se);
1903 update_load_set(&se->load, weight);
1906 account_entity_enqueue(cfs_rq, se);
1909 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1911 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1913 struct task_group *tg;
1914 struct sched_entity *se;
1918 se = tg->se[cpu_of(rq_of(cfs_rq))];
1919 if (!se || throttled_hierarchy(cfs_rq))
1922 if (likely(se->load.weight == tg->shares))
1925 shares = calc_cfs_shares(cfs_rq, tg);
1927 reweight_entity(cfs_rq_of(se), se, shares);
1929 #else /* CONFIG_FAIR_GROUP_SCHED */
1930 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1933 #endif /* CONFIG_FAIR_GROUP_SCHED */
1937 * We choose a half-life close to 1 scheduling period.
1938 * Note: The tables below are dependent on this value.
1940 #define LOAD_AVG_PERIOD 32
1941 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1942 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1944 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1945 static const u32 runnable_avg_yN_inv[] = {
1946 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1947 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1948 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1949 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1950 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1951 0x85aac367, 0x82cd8698,
1955 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1956 * over-estimates when re-combining.
1958 static const u32 runnable_avg_yN_sum[] = {
1959 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1960 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1961 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1966 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1968 static __always_inline u64 decay_load(u64 val, u64 n)
1970 unsigned int local_n;
1974 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1977 /* after bounds checking we can collapse to 32-bit */
1981 * As y^PERIOD = 1/2, we can combine
1982 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1983 * With a look-up table which covers k^n (n<PERIOD)
1985 * To achieve constant time decay_load.
1987 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1988 val >>= local_n / LOAD_AVG_PERIOD;
1989 local_n %= LOAD_AVG_PERIOD;
1992 val *= runnable_avg_yN_inv[local_n];
1993 /* We don't use SRR here since we always want to round down. */
1998 * For updates fully spanning n periods, the contribution to runnable
1999 * average will be: \Sum 1024*y^n
2001 * We can compute this reasonably efficiently by combining:
2002 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
2004 static u32 __compute_runnable_contrib(u64 n)
2008 if (likely(n <= LOAD_AVG_PERIOD))
2009 return runnable_avg_yN_sum[n];
2010 else if (unlikely(n >= LOAD_AVG_MAX_N))
2011 return LOAD_AVG_MAX;
2013 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
2015 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
2016 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
2018 n -= LOAD_AVG_PERIOD;
2019 } while (n > LOAD_AVG_PERIOD);
2021 contrib = decay_load(contrib, n);
2022 return contrib + runnable_avg_yN_sum[n];
2026 * We can represent the historical contribution to runnable average as the
2027 * coefficients of a geometric series. To do this we sub-divide our runnable
2028 * history into segments of approximately 1ms (1024us); label the segment that
2029 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2031 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2033 * (now) (~1ms ago) (~2ms ago)
2035 * Let u_i denote the fraction of p_i that the entity was runnable.
2037 * We then designate the fractions u_i as our co-efficients, yielding the
2038 * following representation of historical load:
2039 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2041 * We choose y based on the with of a reasonably scheduling period, fixing:
2044 * This means that the contribution to load ~32ms ago (u_32) will be weighted
2045 * approximately half as much as the contribution to load within the last ms
2048 * When a period "rolls over" and we have new u_0`, multiplying the previous
2049 * sum again by y is sufficient to update:
2050 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2051 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2053 static __always_inline int __update_entity_runnable_avg(u64 now,
2054 struct sched_avg *sa,
2058 u32 runnable_contrib;
2059 int delta_w, decayed = 0;
2061 delta = now - sa->last_runnable_update;
2063 * This should only happen when time goes backwards, which it
2064 * unfortunately does during sched clock init when we swap over to TSC.
2066 if ((s64)delta < 0) {
2067 sa->last_runnable_update = now;
2072 * Use 1024ns as the unit of measurement since it's a reasonable
2073 * approximation of 1us and fast to compute.
2078 sa->last_runnable_update = now;
2080 /* delta_w is the amount already accumulated against our next period */
2081 delta_w = sa->runnable_avg_period % 1024;
2082 if (delta + delta_w >= 1024) {
2083 /* period roll-over */
2087 * Now that we know we're crossing a period boundary, figure
2088 * out how much from delta we need to complete the current
2089 * period and accrue it.
2091 delta_w = 1024 - delta_w;
2093 sa->runnable_avg_sum += delta_w;
2094 sa->runnable_avg_period += delta_w;
2098 /* Figure out how many additional periods this update spans */
2099 periods = delta / 1024;
2102 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2104 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2107 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2108 runnable_contrib = __compute_runnable_contrib(periods);
2110 sa->runnable_avg_sum += runnable_contrib;
2111 sa->runnable_avg_period += runnable_contrib;
2114 /* Remainder of delta accrued against u_0` */
2116 sa->runnable_avg_sum += delta;
2117 sa->runnable_avg_period += delta;
2122 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2123 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2125 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2126 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2128 decays -= se->avg.decay_count;
2132 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2133 se->avg.decay_count = 0;
2138 #ifdef CONFIG_FAIR_GROUP_SCHED
2139 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2142 struct task_group *tg = cfs_rq->tg;
2145 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2146 tg_contrib -= cfs_rq->tg_load_contrib;
2148 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2149 atomic_long_add(tg_contrib, &tg->load_avg);
2150 cfs_rq->tg_load_contrib += tg_contrib;
2155 * Aggregate cfs_rq runnable averages into an equivalent task_group
2156 * representation for computing load contributions.
2158 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2159 struct cfs_rq *cfs_rq)
2161 struct task_group *tg = cfs_rq->tg;
2164 /* The fraction of a cpu used by this cfs_rq */
2165 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2166 sa->runnable_avg_period + 1);
2167 contrib -= cfs_rq->tg_runnable_contrib;
2169 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2170 atomic_add(contrib, &tg->runnable_avg);
2171 cfs_rq->tg_runnable_contrib += contrib;
2175 static inline void __update_group_entity_contrib(struct sched_entity *se)
2177 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2178 struct task_group *tg = cfs_rq->tg;
2183 contrib = cfs_rq->tg_load_contrib * tg->shares;
2184 se->avg.load_avg_contrib = div_u64(contrib,
2185 atomic_long_read(&tg->load_avg) + 1);
2188 * For group entities we need to compute a correction term in the case
2189 * that they are consuming <1 cpu so that we would contribute the same
2190 * load as a task of equal weight.
2192 * Explicitly co-ordinating this measurement would be expensive, but
2193 * fortunately the sum of each cpus contribution forms a usable
2194 * lower-bound on the true value.
2196 * Consider the aggregate of 2 contributions. Either they are disjoint
2197 * (and the sum represents true value) or they are disjoint and we are
2198 * understating by the aggregate of their overlap.
2200 * Extending this to N cpus, for a given overlap, the maximum amount we
2201 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2202 * cpus that overlap for this interval and w_i is the interval width.
2204 * On a small machine; the first term is well-bounded which bounds the
2205 * total error since w_i is a subset of the period. Whereas on a
2206 * larger machine, while this first term can be larger, if w_i is the
2207 * of consequential size guaranteed to see n_i*w_i quickly converge to
2208 * our upper bound of 1-cpu.
2210 runnable_avg = atomic_read(&tg->runnable_avg);
2211 if (runnable_avg < NICE_0_LOAD) {
2212 se->avg.load_avg_contrib *= runnable_avg;
2213 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2217 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2218 int force_update) {}
2219 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2220 struct cfs_rq *cfs_rq) {}
2221 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2224 static inline void __update_task_entity_contrib(struct sched_entity *se)
2228 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2229 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2230 contrib /= (se->avg.runnable_avg_period + 1);
2231 se->avg.load_avg_contrib = scale_load(contrib);
2234 /* Compute the current contribution to load_avg by se, return any delta */
2235 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2237 long old_contrib = se->avg.load_avg_contrib;
2239 if (entity_is_task(se)) {
2240 __update_task_entity_contrib(se);
2242 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2243 __update_group_entity_contrib(se);
2246 return se->avg.load_avg_contrib - old_contrib;
2249 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2252 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2253 cfs_rq->blocked_load_avg -= load_contrib;
2255 cfs_rq->blocked_load_avg = 0;
2258 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2260 /* Update a sched_entity's runnable average */
2261 static inline void update_entity_load_avg(struct sched_entity *se,
2264 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2269 * For a group entity we need to use their owned cfs_rq_clock_task() in
2270 * case they are the parent of a throttled hierarchy.
2272 if (entity_is_task(se))
2273 now = cfs_rq_clock_task(cfs_rq);
2275 now = cfs_rq_clock_task(group_cfs_rq(se));
2277 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2280 contrib_delta = __update_entity_load_avg_contrib(se);
2286 cfs_rq->runnable_load_avg += contrib_delta;
2288 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2292 * Decay the load contributed by all blocked children and account this so that
2293 * their contribution may appropriately discounted when they wake up.
2295 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2297 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2300 decays = now - cfs_rq->last_decay;
2301 if (!decays && !force_update)
2304 if (atomic_long_read(&cfs_rq->removed_load)) {
2305 unsigned long removed_load;
2306 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2307 subtract_blocked_load_contrib(cfs_rq, removed_load);
2311 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2313 atomic64_add(decays, &cfs_rq->decay_counter);
2314 cfs_rq->last_decay = now;
2317 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2320 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2322 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2323 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2326 /* Add the load generated by se into cfs_rq's child load-average */
2327 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2328 struct sched_entity *se,
2332 * We track migrations using entity decay_count <= 0, on a wake-up
2333 * migration we use a negative decay count to track the remote decays
2334 * accumulated while sleeping.
2336 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2337 * are seen by enqueue_entity_load_avg() as a migration with an already
2338 * constructed load_avg_contrib.
2340 if (unlikely(se->avg.decay_count <= 0)) {
2341 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2342 if (se->avg.decay_count) {
2344 * In a wake-up migration we have to approximate the
2345 * time sleeping. This is because we can't synchronize
2346 * clock_task between the two cpus, and it is not
2347 * guaranteed to be read-safe. Instead, we can
2348 * approximate this using our carried decays, which are
2349 * explicitly atomically readable.
2351 se->avg.last_runnable_update -= (-se->avg.decay_count)
2353 update_entity_load_avg(se, 0);
2354 /* Indicate that we're now synchronized and on-rq */
2355 se->avg.decay_count = 0;
2360 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2361 * would have made count negative); we must be careful to avoid
2362 * double-accounting blocked time after synchronizing decays.
2364 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2368 /* migrated tasks did not contribute to our blocked load */
2370 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2371 update_entity_load_avg(se, 0);
2374 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2375 /* we force update consideration on load-balancer moves */
2376 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2380 * Remove se's load from this cfs_rq child load-average, if the entity is
2381 * transitioning to a blocked state we track its projected decay using
2384 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2385 struct sched_entity *se,
2388 update_entity_load_avg(se, 1);
2389 /* we force update consideration on load-balancer moves */
2390 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2392 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2394 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2395 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2396 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2400 * Update the rq's load with the elapsed running time before entering
2401 * idle. if the last scheduled task is not a CFS task, idle_enter will
2402 * be the only way to update the runnable statistic.
2404 void idle_enter_fair(struct rq *this_rq)
2406 update_rq_runnable_avg(this_rq, 1);
2410 * Update the rq's load with the elapsed idle time before a task is
2411 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2412 * be the only way to update the runnable statistic.
2414 void idle_exit_fair(struct rq *this_rq)
2416 update_rq_runnable_avg(this_rq, 0);
2420 static inline void update_entity_load_avg(struct sched_entity *se,
2421 int update_cfs_rq) {}
2422 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2423 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2424 struct sched_entity *se,
2426 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2427 struct sched_entity *se,
2429 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2430 int force_update) {}
2433 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2435 #ifdef CONFIG_SCHEDSTATS
2436 struct task_struct *tsk = NULL;
2438 if (entity_is_task(se))
2441 if (se->statistics.sleep_start) {
2442 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2447 if (unlikely(delta > se->statistics.sleep_max))
2448 se->statistics.sleep_max = delta;
2450 se->statistics.sleep_start = 0;
2451 se->statistics.sum_sleep_runtime += delta;
2454 account_scheduler_latency(tsk, delta >> 10, 1);
2455 trace_sched_stat_sleep(tsk, delta);
2458 if (se->statistics.block_start) {
2459 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2464 if (unlikely(delta > se->statistics.block_max))
2465 se->statistics.block_max = delta;
2467 se->statistics.block_start = 0;
2468 se->statistics.sum_sleep_runtime += delta;
2471 if (tsk->in_iowait) {
2472 se->statistics.iowait_sum += delta;
2473 se->statistics.iowait_count++;
2474 trace_sched_stat_iowait(tsk, delta);
2477 trace_sched_stat_blocked(tsk, delta);
2480 * Blocking time is in units of nanosecs, so shift by
2481 * 20 to get a milliseconds-range estimation of the
2482 * amount of time that the task spent sleeping:
2484 if (unlikely(prof_on == SLEEP_PROFILING)) {
2485 profile_hits(SLEEP_PROFILING,
2486 (void *)get_wchan(tsk),
2489 account_scheduler_latency(tsk, delta >> 10, 0);
2495 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2497 #ifdef CONFIG_SCHED_DEBUG
2498 s64 d = se->vruntime - cfs_rq->min_vruntime;
2503 if (d > 3*sysctl_sched_latency)
2504 schedstat_inc(cfs_rq, nr_spread_over);
2509 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2511 u64 vruntime = cfs_rq->min_vruntime;
2514 * The 'current' period is already promised to the current tasks,
2515 * however the extra weight of the new task will slow them down a
2516 * little, place the new task so that it fits in the slot that
2517 * stays open at the end.
2519 if (initial && sched_feat(START_DEBIT))
2520 vruntime += sched_vslice(cfs_rq, se);
2522 /* sleeps up to a single latency don't count. */
2524 unsigned long thresh = sysctl_sched_latency;
2527 * Halve their sleep time's effect, to allow
2528 * for a gentler effect of sleepers:
2530 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2536 /* ensure we never gain time by being placed backwards. */
2537 se->vruntime = max_vruntime(se->vruntime, vruntime);
2540 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2543 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2546 * Update the normalized vruntime before updating min_vruntime
2547 * through calling update_curr().
2549 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2550 se->vruntime += cfs_rq->min_vruntime;
2553 * Update run-time statistics of the 'current'.
2555 update_curr(cfs_rq);
2556 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2557 account_entity_enqueue(cfs_rq, se);
2558 update_cfs_shares(cfs_rq);
2560 if (flags & ENQUEUE_WAKEUP) {
2561 place_entity(cfs_rq, se, 0);
2562 enqueue_sleeper(cfs_rq, se);
2565 update_stats_enqueue(cfs_rq, se);
2566 check_spread(cfs_rq, se);
2567 if (se != cfs_rq->curr)
2568 __enqueue_entity(cfs_rq, se);
2571 if (cfs_rq->nr_running == 1) {
2572 list_add_leaf_cfs_rq(cfs_rq);
2573 check_enqueue_throttle(cfs_rq);
2577 static void __clear_buddies_last(struct sched_entity *se)
2579 for_each_sched_entity(se) {
2580 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2581 if (cfs_rq->last == se)
2582 cfs_rq->last = NULL;
2588 static void __clear_buddies_next(struct sched_entity *se)
2590 for_each_sched_entity(se) {
2591 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2592 if (cfs_rq->next == se)
2593 cfs_rq->next = NULL;
2599 static void __clear_buddies_skip(struct sched_entity *se)
2601 for_each_sched_entity(se) {
2602 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2603 if (cfs_rq->skip == se)
2604 cfs_rq->skip = NULL;
2610 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2612 if (cfs_rq->last == se)
2613 __clear_buddies_last(se);
2615 if (cfs_rq->next == se)
2616 __clear_buddies_next(se);
2618 if (cfs_rq->skip == se)
2619 __clear_buddies_skip(se);
2622 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2625 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2628 * Update run-time statistics of the 'current'.
2630 update_curr(cfs_rq);
2631 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2633 update_stats_dequeue(cfs_rq, se);
2634 if (flags & DEQUEUE_SLEEP) {
2635 #ifdef CONFIG_SCHEDSTATS
2636 if (entity_is_task(se)) {
2637 struct task_struct *tsk = task_of(se);
2639 if (tsk->state & TASK_INTERRUPTIBLE)
2640 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2641 if (tsk->state & TASK_UNINTERRUPTIBLE)
2642 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2647 clear_buddies(cfs_rq, se);
2649 if (se != cfs_rq->curr)
2650 __dequeue_entity(cfs_rq, se);
2652 account_entity_dequeue(cfs_rq, se);
2655 * Normalize the entity after updating the min_vruntime because the
2656 * update can refer to the ->curr item and we need to reflect this
2657 * movement in our normalized position.
2659 if (!(flags & DEQUEUE_SLEEP))
2660 se->vruntime -= cfs_rq->min_vruntime;
2662 /* return excess runtime on last dequeue */
2663 return_cfs_rq_runtime(cfs_rq);
2665 update_min_vruntime(cfs_rq);
2666 update_cfs_shares(cfs_rq);
2670 * Preempt the current task with a newly woken task if needed:
2673 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2675 unsigned long ideal_runtime, delta_exec;
2676 struct sched_entity *se;
2679 ideal_runtime = sched_slice(cfs_rq, curr);
2680 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2681 if (delta_exec > ideal_runtime) {
2682 resched_task(rq_of(cfs_rq)->curr);
2684 * The current task ran long enough, ensure it doesn't get
2685 * re-elected due to buddy favours.
2687 clear_buddies(cfs_rq, curr);
2692 * Ensure that a task that missed wakeup preemption by a
2693 * narrow margin doesn't have to wait for a full slice.
2694 * This also mitigates buddy induced latencies under load.
2696 if (delta_exec < sysctl_sched_min_granularity)
2699 se = __pick_first_entity(cfs_rq);
2700 delta = curr->vruntime - se->vruntime;
2705 if (delta > ideal_runtime)
2706 resched_task(rq_of(cfs_rq)->curr);
2710 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2712 /* 'current' is not kept within the tree. */
2715 * Any task has to be enqueued before it get to execute on
2716 * a CPU. So account for the time it spent waiting on the
2719 update_stats_wait_end(cfs_rq, se);
2720 __dequeue_entity(cfs_rq, se);
2723 update_stats_curr_start(cfs_rq, se);
2725 #ifdef CONFIG_SCHEDSTATS
2727 * Track our maximum slice length, if the CPU's load is at
2728 * least twice that of our own weight (i.e. dont track it
2729 * when there are only lesser-weight tasks around):
2731 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2732 se->statistics.slice_max = max(se->statistics.slice_max,
2733 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2736 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2740 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2743 * Pick the next process, keeping these things in mind, in this order:
2744 * 1) keep things fair between processes/task groups
2745 * 2) pick the "next" process, since someone really wants that to run
2746 * 3) pick the "last" process, for cache locality
2747 * 4) do not run the "skip" process, if something else is available
2749 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2751 struct sched_entity *se = __pick_first_entity(cfs_rq);
2752 struct sched_entity *left = se;
2755 * Avoid running the skip buddy, if running something else can
2756 * be done without getting too unfair.
2758 if (cfs_rq->skip == se) {
2759 struct sched_entity *second = __pick_next_entity(se);
2760 if (second && wakeup_preempt_entity(second, left) < 1)
2765 * Prefer last buddy, try to return the CPU to a preempted task.
2767 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2771 * Someone really wants this to run. If it's not unfair, run it.
2773 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2776 clear_buddies(cfs_rq, se);
2781 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2783 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2786 * If still on the runqueue then deactivate_task()
2787 * was not called and update_curr() has to be done:
2790 update_curr(cfs_rq);
2792 /* throttle cfs_rqs exceeding runtime */
2793 check_cfs_rq_runtime(cfs_rq);
2795 check_spread(cfs_rq, prev);
2797 update_stats_wait_start(cfs_rq, prev);
2798 /* Put 'current' back into the tree. */
2799 __enqueue_entity(cfs_rq, prev);
2800 /* in !on_rq case, update occurred at dequeue */
2801 update_entity_load_avg(prev, 1);
2803 cfs_rq->curr = NULL;
2807 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2810 * Update run-time statistics of the 'current'.
2812 update_curr(cfs_rq);
2815 * Ensure that runnable average is periodically updated.
2817 update_entity_load_avg(curr, 1);
2818 update_cfs_rq_blocked_load(cfs_rq, 1);
2819 update_cfs_shares(cfs_rq);
2821 #ifdef CONFIG_SCHED_HRTICK
2823 * queued ticks are scheduled to match the slice, so don't bother
2824 * validating it and just reschedule.
2827 resched_task(rq_of(cfs_rq)->curr);
2831 * don't let the period tick interfere with the hrtick preemption
2833 if (!sched_feat(DOUBLE_TICK) &&
2834 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2838 if (cfs_rq->nr_running > 1)
2839 check_preempt_tick(cfs_rq, curr);
2843 /**************************************************
2844 * CFS bandwidth control machinery
2847 #ifdef CONFIG_CFS_BANDWIDTH
2849 #ifdef HAVE_JUMP_LABEL
2850 static struct static_key __cfs_bandwidth_used;
2852 static inline bool cfs_bandwidth_used(void)
2854 return static_key_false(&__cfs_bandwidth_used);
2857 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2859 /* only need to count groups transitioning between enabled/!enabled */
2860 if (enabled && !was_enabled)
2861 static_key_slow_inc(&__cfs_bandwidth_used);
2862 else if (!enabled && was_enabled)
2863 static_key_slow_dec(&__cfs_bandwidth_used);
2865 #else /* HAVE_JUMP_LABEL */
2866 static bool cfs_bandwidth_used(void)
2871 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2872 #endif /* HAVE_JUMP_LABEL */
2875 * default period for cfs group bandwidth.
2876 * default: 0.1s, units: nanoseconds
2878 static inline u64 default_cfs_period(void)
2880 return 100000000ULL;
2883 static inline u64 sched_cfs_bandwidth_slice(void)
2885 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2889 * Replenish runtime according to assigned quota and update expiration time.
2890 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2891 * additional synchronization around rq->lock.
2893 * requires cfs_b->lock
2895 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2899 if (cfs_b->quota == RUNTIME_INF)
2902 now = sched_clock_cpu(smp_processor_id());
2903 cfs_b->runtime = cfs_b->quota;
2904 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2907 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2909 return &tg->cfs_bandwidth;
2912 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2913 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2915 if (unlikely(cfs_rq->throttle_count))
2916 return cfs_rq->throttled_clock_task;
2918 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2921 /* returns 0 on failure to allocate runtime */
2922 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2924 struct task_group *tg = cfs_rq->tg;
2925 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2926 u64 amount = 0, min_amount, expires;
2928 /* note: this is a positive sum as runtime_remaining <= 0 */
2929 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2931 raw_spin_lock(&cfs_b->lock);
2932 if (cfs_b->quota == RUNTIME_INF)
2933 amount = min_amount;
2936 * If the bandwidth pool has become inactive, then at least one
2937 * period must have elapsed since the last consumption.
2938 * Refresh the global state and ensure bandwidth timer becomes
2941 if (!cfs_b->timer_active) {
2942 __refill_cfs_bandwidth_runtime(cfs_b);
2943 __start_cfs_bandwidth(cfs_b);
2946 if (cfs_b->runtime > 0) {
2947 amount = min(cfs_b->runtime, min_amount);
2948 cfs_b->runtime -= amount;
2952 expires = cfs_b->runtime_expires;
2953 raw_spin_unlock(&cfs_b->lock);
2955 cfs_rq->runtime_remaining += amount;
2957 * we may have advanced our local expiration to account for allowed
2958 * spread between our sched_clock and the one on which runtime was
2961 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2962 cfs_rq->runtime_expires = expires;
2964 return cfs_rq->runtime_remaining > 0;
2968 * Note: This depends on the synchronization provided by sched_clock and the
2969 * fact that rq->clock snapshots this value.
2971 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2973 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2975 /* if the deadline is ahead of our clock, nothing to do */
2976 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2979 if (cfs_rq->runtime_remaining < 0)
2983 * If the local deadline has passed we have to consider the
2984 * possibility that our sched_clock is 'fast' and the global deadline
2985 * has not truly expired.
2987 * Fortunately we can check determine whether this the case by checking
2988 * whether the global deadline has advanced.
2991 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2992 /* extend local deadline, drift is bounded above by 2 ticks */
2993 cfs_rq->runtime_expires += TICK_NSEC;
2995 /* global deadline is ahead, expiration has passed */
2996 cfs_rq->runtime_remaining = 0;
3000 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3001 unsigned long delta_exec)
3003 /* dock delta_exec before expiring quota (as it could span periods) */
3004 cfs_rq->runtime_remaining -= delta_exec;
3005 expire_cfs_rq_runtime(cfs_rq);
3007 if (likely(cfs_rq->runtime_remaining > 0))
3011 * if we're unable to extend our runtime we resched so that the active
3012 * hierarchy can be throttled
3014 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3015 resched_task(rq_of(cfs_rq)->curr);
3018 static __always_inline
3019 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
3021 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3024 __account_cfs_rq_runtime(cfs_rq, delta_exec);
3027 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3029 return cfs_bandwidth_used() && cfs_rq->throttled;
3032 /* check whether cfs_rq, or any parent, is throttled */
3033 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3035 return cfs_bandwidth_used() && cfs_rq->throttle_count;
3039 * Ensure that neither of the group entities corresponding to src_cpu or
3040 * dest_cpu are members of a throttled hierarchy when performing group
3041 * load-balance operations.
3043 static inline int throttled_lb_pair(struct task_group *tg,
3044 int src_cpu, int dest_cpu)
3046 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3048 src_cfs_rq = tg->cfs_rq[src_cpu];
3049 dest_cfs_rq = tg->cfs_rq[dest_cpu];
3051 return throttled_hierarchy(src_cfs_rq) ||
3052 throttled_hierarchy(dest_cfs_rq);
3055 /* updated child weight may affect parent so we have to do this bottom up */
3056 static int tg_unthrottle_up(struct task_group *tg, void *data)
3058 struct rq *rq = data;
3059 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3061 cfs_rq->throttle_count--;
3063 if (!cfs_rq->throttle_count) {
3064 /* adjust cfs_rq_clock_task() */
3065 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3066 cfs_rq->throttled_clock_task;
3073 static int tg_throttle_down(struct task_group *tg, void *data)
3075 struct rq *rq = data;
3076 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3078 /* group is entering throttled state, stop time */
3079 if (!cfs_rq->throttle_count)
3080 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3081 cfs_rq->throttle_count++;
3086 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3088 struct rq *rq = rq_of(cfs_rq);
3089 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3090 struct sched_entity *se;
3091 long task_delta, dequeue = 1;
3093 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3095 /* freeze hierarchy runnable averages while throttled */
3097 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3100 task_delta = cfs_rq->h_nr_running;
3101 for_each_sched_entity(se) {
3102 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3103 /* throttled entity or throttle-on-deactivate */
3108 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3109 qcfs_rq->h_nr_running -= task_delta;
3111 if (qcfs_rq->load.weight)
3116 rq->nr_running -= task_delta;
3118 cfs_rq->throttled = 1;
3119 cfs_rq->throttled_clock = rq_clock(rq);
3120 raw_spin_lock(&cfs_b->lock);
3121 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3122 raw_spin_unlock(&cfs_b->lock);
3125 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3127 struct rq *rq = rq_of(cfs_rq);
3128 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3129 struct sched_entity *se;
3133 se = cfs_rq->tg->se[cpu_of(rq)];
3135 cfs_rq->throttled = 0;
3137 update_rq_clock(rq);
3139 raw_spin_lock(&cfs_b->lock);
3140 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3141 list_del_rcu(&cfs_rq->throttled_list);
3142 raw_spin_unlock(&cfs_b->lock);
3144 /* update hierarchical throttle state */
3145 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3147 if (!cfs_rq->load.weight)
3150 task_delta = cfs_rq->h_nr_running;
3151 for_each_sched_entity(se) {
3155 cfs_rq = cfs_rq_of(se);
3157 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3158 cfs_rq->h_nr_running += task_delta;
3160 if (cfs_rq_throttled(cfs_rq))
3165 rq->nr_running += task_delta;
3167 /* determine whether we need to wake up potentially idle cpu */
3168 if (rq->curr == rq->idle && rq->cfs.nr_running)
3169 resched_task(rq->curr);
3172 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3173 u64 remaining, u64 expires)
3175 struct cfs_rq *cfs_rq;
3176 u64 runtime = remaining;
3179 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3181 struct rq *rq = rq_of(cfs_rq);
3183 raw_spin_lock(&rq->lock);
3184 if (!cfs_rq_throttled(cfs_rq))
3187 runtime = -cfs_rq->runtime_remaining + 1;
3188 if (runtime > remaining)
3189 runtime = remaining;
3190 remaining -= runtime;
3192 cfs_rq->runtime_remaining += runtime;
3193 cfs_rq->runtime_expires = expires;
3195 /* we check whether we're throttled above */
3196 if (cfs_rq->runtime_remaining > 0)
3197 unthrottle_cfs_rq(cfs_rq);
3200 raw_spin_unlock(&rq->lock);
3211 * Responsible for refilling a task_group's bandwidth and unthrottling its
3212 * cfs_rqs as appropriate. If there has been no activity within the last
3213 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3214 * used to track this state.
3216 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3218 u64 runtime, runtime_expires;
3219 int idle = 1, throttled;
3221 raw_spin_lock(&cfs_b->lock);
3222 /* no need to continue the timer with no bandwidth constraint */
3223 if (cfs_b->quota == RUNTIME_INF)
3226 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3227 /* idle depends on !throttled (for the case of a large deficit) */
3228 idle = cfs_b->idle && !throttled;
3229 cfs_b->nr_periods += overrun;
3231 /* if we're going inactive then everything else can be deferred */
3235 __refill_cfs_bandwidth_runtime(cfs_b);
3238 /* mark as potentially idle for the upcoming period */
3243 /* account preceding periods in which throttling occurred */
3244 cfs_b->nr_throttled += overrun;
3247 * There are throttled entities so we must first use the new bandwidth
3248 * to unthrottle them before making it generally available. This
3249 * ensures that all existing debts will be paid before a new cfs_rq is
3252 runtime = cfs_b->runtime;
3253 runtime_expires = cfs_b->runtime_expires;
3257 * This check is repeated as we are holding onto the new bandwidth
3258 * while we unthrottle. This can potentially race with an unthrottled
3259 * group trying to acquire new bandwidth from the global pool.
3261 while (throttled && runtime > 0) {
3262 raw_spin_unlock(&cfs_b->lock);
3263 /* we can't nest cfs_b->lock while distributing bandwidth */
3264 runtime = distribute_cfs_runtime(cfs_b, runtime,
3266 raw_spin_lock(&cfs_b->lock);
3268 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3271 /* return (any) remaining runtime */
3272 cfs_b->runtime = runtime;
3274 * While we are ensured activity in the period following an
3275 * unthrottle, this also covers the case in which the new bandwidth is
3276 * insufficient to cover the existing bandwidth deficit. (Forcing the
3277 * timer to remain active while there are any throttled entities.)
3282 cfs_b->timer_active = 0;
3283 raw_spin_unlock(&cfs_b->lock);
3288 /* a cfs_rq won't donate quota below this amount */
3289 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3290 /* minimum remaining period time to redistribute slack quota */
3291 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3292 /* how long we wait to gather additional slack before distributing */
3293 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3295 /* are we near the end of the current quota period? */
3296 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3298 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3301 /* if the call-back is running a quota refresh is already occurring */
3302 if (hrtimer_callback_running(refresh_timer))
3305 /* is a quota refresh about to occur? */
3306 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3307 if (remaining < min_expire)
3313 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3315 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3317 /* if there's a quota refresh soon don't bother with slack */
3318 if (runtime_refresh_within(cfs_b, min_left))
3321 start_bandwidth_timer(&cfs_b->slack_timer,
3322 ns_to_ktime(cfs_bandwidth_slack_period));
3325 /* we know any runtime found here is valid as update_curr() precedes return */
3326 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3328 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3329 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3331 if (slack_runtime <= 0)
3334 raw_spin_lock(&cfs_b->lock);
3335 if (cfs_b->quota != RUNTIME_INF &&
3336 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3337 cfs_b->runtime += slack_runtime;
3339 /* we are under rq->lock, defer unthrottling using a timer */
3340 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3341 !list_empty(&cfs_b->throttled_cfs_rq))
3342 start_cfs_slack_bandwidth(cfs_b);
3344 raw_spin_unlock(&cfs_b->lock);
3346 /* even if it's not valid for return we don't want to try again */
3347 cfs_rq->runtime_remaining -= slack_runtime;
3350 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3352 if (!cfs_bandwidth_used())
3355 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3358 __return_cfs_rq_runtime(cfs_rq);
3362 * This is done with a timer (instead of inline with bandwidth return) since
3363 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3365 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3367 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3370 /* confirm we're still not at a refresh boundary */
3371 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3374 raw_spin_lock(&cfs_b->lock);
3375 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3376 runtime = cfs_b->runtime;
3379 expires = cfs_b->runtime_expires;
3380 raw_spin_unlock(&cfs_b->lock);
3385 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3387 raw_spin_lock(&cfs_b->lock);
3388 if (expires == cfs_b->runtime_expires)
3389 cfs_b->runtime = runtime;
3390 raw_spin_unlock(&cfs_b->lock);
3394 * When a group wakes up we want to make sure that its quota is not already
3395 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3396 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3398 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3400 if (!cfs_bandwidth_used())
3403 /* an active group must be handled by the update_curr()->put() path */
3404 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3407 /* ensure the group is not already throttled */
3408 if (cfs_rq_throttled(cfs_rq))
3411 /* update runtime allocation */
3412 account_cfs_rq_runtime(cfs_rq, 0);
3413 if (cfs_rq->runtime_remaining <= 0)
3414 throttle_cfs_rq(cfs_rq);
3417 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3418 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3420 if (!cfs_bandwidth_used())
3423 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3427 * it's possible for a throttled entity to be forced into a running
3428 * state (e.g. set_curr_task), in this case we're finished.
3430 if (cfs_rq_throttled(cfs_rq))
3433 throttle_cfs_rq(cfs_rq);
3436 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3438 struct cfs_bandwidth *cfs_b =
3439 container_of(timer, struct cfs_bandwidth, slack_timer);
3440 do_sched_cfs_slack_timer(cfs_b);
3442 return HRTIMER_NORESTART;
3445 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3447 struct cfs_bandwidth *cfs_b =
3448 container_of(timer, struct cfs_bandwidth, period_timer);
3454 now = hrtimer_cb_get_time(timer);
3455 overrun = hrtimer_forward(timer, now, cfs_b->period);
3460 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3463 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3466 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3468 raw_spin_lock_init(&cfs_b->lock);
3470 cfs_b->quota = RUNTIME_INF;
3471 cfs_b->period = ns_to_ktime(default_cfs_period());
3473 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3474 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3475 cfs_b->period_timer.function = sched_cfs_period_timer;
3476 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3477 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3480 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3482 cfs_rq->runtime_enabled = 0;
3483 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3486 /* requires cfs_b->lock, may release to reprogram timer */
3487 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3490 * The timer may be active because we're trying to set a new bandwidth
3491 * period or because we're racing with the tear-down path
3492 * (timer_active==0 becomes visible before the hrtimer call-back
3493 * terminates). In either case we ensure that it's re-programmed
3495 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3496 raw_spin_unlock(&cfs_b->lock);
3497 /* ensure cfs_b->lock is available while we wait */
3498 hrtimer_cancel(&cfs_b->period_timer);
3500 raw_spin_lock(&cfs_b->lock);
3501 /* if someone else restarted the timer then we're done */
3502 if (cfs_b->timer_active)
3506 cfs_b->timer_active = 1;
3507 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3510 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3512 hrtimer_cancel(&cfs_b->period_timer);
3513 hrtimer_cancel(&cfs_b->slack_timer);
3516 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3518 struct cfs_rq *cfs_rq;
3520 for_each_leaf_cfs_rq(rq, cfs_rq) {
3521 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3523 if (!cfs_rq->runtime_enabled)
3527 * clock_task is not advancing so we just need to make sure
3528 * there's some valid quota amount
3530 cfs_rq->runtime_remaining = cfs_b->quota;
3531 if (cfs_rq_throttled(cfs_rq))
3532 unthrottle_cfs_rq(cfs_rq);
3536 #else /* CONFIG_CFS_BANDWIDTH */
3537 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3539 return rq_clock_task(rq_of(cfs_rq));
3542 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3543 unsigned long delta_exec) {}
3544 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3545 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3546 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3548 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3553 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3558 static inline int throttled_lb_pair(struct task_group *tg,
3559 int src_cpu, int dest_cpu)
3564 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3566 #ifdef CONFIG_FAIR_GROUP_SCHED
3567 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3570 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3574 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3575 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3577 #endif /* CONFIG_CFS_BANDWIDTH */
3579 /**************************************************
3580 * CFS operations on tasks:
3583 #ifdef CONFIG_SCHED_HRTICK
3584 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3586 struct sched_entity *se = &p->se;
3587 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3589 WARN_ON(task_rq(p) != rq);
3591 if (cfs_rq->nr_running > 1) {
3592 u64 slice = sched_slice(cfs_rq, se);
3593 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3594 s64 delta = slice - ran;
3603 * Don't schedule slices shorter than 10000ns, that just
3604 * doesn't make sense. Rely on vruntime for fairness.
3607 delta = max_t(s64, 10000LL, delta);
3609 hrtick_start(rq, delta);
3614 * called from enqueue/dequeue and updates the hrtick when the
3615 * current task is from our class and nr_running is low enough
3618 static void hrtick_update(struct rq *rq)
3620 struct task_struct *curr = rq->curr;
3622 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3625 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3626 hrtick_start_fair(rq, curr);
3628 #else /* !CONFIG_SCHED_HRTICK */
3630 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3634 static inline void hrtick_update(struct rq *rq)
3640 * The enqueue_task method is called before nr_running is
3641 * increased. Here we update the fair scheduling stats and
3642 * then put the task into the rbtree:
3645 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3647 struct cfs_rq *cfs_rq;
3648 struct sched_entity *se = &p->se;
3650 for_each_sched_entity(se) {
3653 cfs_rq = cfs_rq_of(se);
3654 enqueue_entity(cfs_rq, se, flags);
3657 * end evaluation on encountering a throttled cfs_rq
3659 * note: in the case of encountering a throttled cfs_rq we will
3660 * post the final h_nr_running increment below.
3662 if (cfs_rq_throttled(cfs_rq))
3664 cfs_rq->h_nr_running++;
3666 flags = ENQUEUE_WAKEUP;
3669 for_each_sched_entity(se) {
3670 cfs_rq = cfs_rq_of(se);
3671 cfs_rq->h_nr_running++;
3673 if (cfs_rq_throttled(cfs_rq))
3676 update_cfs_shares(cfs_rq);
3677 update_entity_load_avg(se, 1);
3681 update_rq_runnable_avg(rq, rq->nr_running);
3687 static void set_next_buddy(struct sched_entity *se);
3690 * The dequeue_task method is called before nr_running is
3691 * decreased. We remove the task from the rbtree and
3692 * update the fair scheduling stats:
3694 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3696 struct cfs_rq *cfs_rq;
3697 struct sched_entity *se = &p->se;
3698 int task_sleep = flags & DEQUEUE_SLEEP;
3700 for_each_sched_entity(se) {
3701 cfs_rq = cfs_rq_of(se);
3702 dequeue_entity(cfs_rq, se, flags);
3705 * end evaluation on encountering a throttled cfs_rq
3707 * note: in the case of encountering a throttled cfs_rq we will
3708 * post the final h_nr_running decrement below.
3710 if (cfs_rq_throttled(cfs_rq))
3712 cfs_rq->h_nr_running--;
3714 /* Don't dequeue parent if it has other entities besides us */
3715 if (cfs_rq->load.weight) {
3717 * Bias pick_next to pick a task from this cfs_rq, as
3718 * p is sleeping when it is within its sched_slice.
3720 if (task_sleep && parent_entity(se))
3721 set_next_buddy(parent_entity(se));
3723 /* avoid re-evaluating load for this entity */
3724 se = parent_entity(se);
3727 flags |= DEQUEUE_SLEEP;
3730 for_each_sched_entity(se) {
3731 cfs_rq = cfs_rq_of(se);
3732 cfs_rq->h_nr_running--;
3734 if (cfs_rq_throttled(cfs_rq))
3737 update_cfs_shares(cfs_rq);
3738 update_entity_load_avg(se, 1);
3743 update_rq_runnable_avg(rq, 1);
3749 /* Used instead of source_load when we know the type == 0 */
3750 static unsigned long weighted_cpuload(const int cpu)
3752 return cpu_rq(cpu)->cfs.runnable_load_avg;
3756 * Return a low guess at the load of a migration-source cpu weighted
3757 * according to the scheduling class and "nice" value.
3759 * We want to under-estimate the load of migration sources, to
3760 * balance conservatively.
3762 static unsigned long source_load(int cpu, int type)
3764 struct rq *rq = cpu_rq(cpu);
3765 unsigned long total = weighted_cpuload(cpu);
3767 if (type == 0 || !sched_feat(LB_BIAS))
3770 return min(rq->cpu_load[type-1], total);
3774 * Return a high guess at the load of a migration-target cpu weighted
3775 * according to the scheduling class and "nice" value.
3777 static unsigned long target_load(int cpu, int type)
3779 struct rq *rq = cpu_rq(cpu);
3780 unsigned long total = weighted_cpuload(cpu);
3782 if (type == 0 || !sched_feat(LB_BIAS))
3785 return max(rq->cpu_load[type-1], total);
3788 static unsigned long power_of(int cpu)
3790 return cpu_rq(cpu)->cpu_power;
3793 static unsigned long cpu_avg_load_per_task(int cpu)
3795 struct rq *rq = cpu_rq(cpu);
3796 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3797 unsigned long load_avg = rq->cfs.runnable_load_avg;
3800 return load_avg / nr_running;
3805 static void record_wakee(struct task_struct *p)
3808 * Rough decay (wiping) for cost saving, don't worry
3809 * about the boundary, really active task won't care
3812 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3813 current->wakee_flips = 0;
3814 current->wakee_flip_decay_ts = jiffies;
3817 if (current->last_wakee != p) {
3818 current->last_wakee = p;
3819 current->wakee_flips++;
3823 static void task_waking_fair(struct task_struct *p)
3825 struct sched_entity *se = &p->se;
3826 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3829 #ifndef CONFIG_64BIT
3830 u64 min_vruntime_copy;
3833 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3835 min_vruntime = cfs_rq->min_vruntime;
3836 } while (min_vruntime != min_vruntime_copy);
3838 min_vruntime = cfs_rq->min_vruntime;
3841 se->vruntime -= min_vruntime;
3845 #ifdef CONFIG_FAIR_GROUP_SCHED
3847 * effective_load() calculates the load change as seen from the root_task_group
3849 * Adding load to a group doesn't make a group heavier, but can cause movement
3850 * of group shares between cpus. Assuming the shares were perfectly aligned one
3851 * can calculate the shift in shares.
3853 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3854 * on this @cpu and results in a total addition (subtraction) of @wg to the
3855 * total group weight.
3857 * Given a runqueue weight distribution (rw_i) we can compute a shares
3858 * distribution (s_i) using:
3860 * s_i = rw_i / \Sum rw_j (1)
3862 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3863 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3864 * shares distribution (s_i):
3866 * rw_i = { 2, 4, 1, 0 }
3867 * s_i = { 2/7, 4/7, 1/7, 0 }
3869 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3870 * task used to run on and the CPU the waker is running on), we need to
3871 * compute the effect of waking a task on either CPU and, in case of a sync
3872 * wakeup, compute the effect of the current task going to sleep.
3874 * So for a change of @wl to the local @cpu with an overall group weight change
3875 * of @wl we can compute the new shares distribution (s'_i) using:
3877 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3879 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3880 * differences in waking a task to CPU 0. The additional task changes the
3881 * weight and shares distributions like:
3883 * rw'_i = { 3, 4, 1, 0 }
3884 * s'_i = { 3/8, 4/8, 1/8, 0 }
3886 * We can then compute the difference in effective weight by using:
3888 * dw_i = S * (s'_i - s_i) (3)
3890 * Where 'S' is the group weight as seen by its parent.
3892 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3893 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3894 * 4/7) times the weight of the group.
3896 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3898 struct sched_entity *se = tg->se[cpu];
3900 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3903 for_each_sched_entity(se) {
3909 * W = @wg + \Sum rw_j
3911 W = wg + calc_tg_weight(tg, se->my_q);
3916 w = se->my_q->load.weight + wl;
3919 * wl = S * s'_i; see (2)
3922 wl = (w * tg->shares) / W;
3927 * Per the above, wl is the new se->load.weight value; since
3928 * those are clipped to [MIN_SHARES, ...) do so now. See
3929 * calc_cfs_shares().
3931 if (wl < MIN_SHARES)
3935 * wl = dw_i = S * (s'_i - s_i); see (3)
3937 wl -= se->load.weight;
3940 * Recursively apply this logic to all parent groups to compute
3941 * the final effective load change on the root group. Since
3942 * only the @tg group gets extra weight, all parent groups can
3943 * only redistribute existing shares. @wl is the shift in shares
3944 * resulting from this level per the above.
3953 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3960 static int wake_wide(struct task_struct *p)
3962 int factor = this_cpu_read(sd_llc_size);
3965 * Yeah, it's the switching-frequency, could means many wakee or
3966 * rapidly switch, use factor here will just help to automatically
3967 * adjust the loose-degree, so bigger node will lead to more pull.
3969 if (p->wakee_flips > factor) {
3971 * wakee is somewhat hot, it needs certain amount of cpu
3972 * resource, so if waker is far more hot, prefer to leave
3975 if (current->wakee_flips > (factor * p->wakee_flips))
3982 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3984 s64 this_load, load;
3985 int idx, this_cpu, prev_cpu;
3986 unsigned long tl_per_task;
3987 struct task_group *tg;
3988 unsigned long weight;
3992 * If we wake multiple tasks be careful to not bounce
3993 * ourselves around too much.
3999 this_cpu = smp_processor_id();
4000 prev_cpu = task_cpu(p);
4001 load = source_load(prev_cpu, idx);
4002 this_load = target_load(this_cpu, idx);
4005 * If sync wakeup then subtract the (maximum possible)
4006 * effect of the currently running task from the load
4007 * of the current CPU:
4010 tg = task_group(current);
4011 weight = current->se.load.weight;
4013 this_load += effective_load(tg, this_cpu, -weight, -weight);
4014 load += effective_load(tg, prev_cpu, 0, -weight);
4018 weight = p->se.load.weight;
4021 * In low-load situations, where prev_cpu is idle and this_cpu is idle
4022 * due to the sync cause above having dropped this_load to 0, we'll
4023 * always have an imbalance, but there's really nothing you can do
4024 * about that, so that's good too.
4026 * Otherwise check if either cpus are near enough in load to allow this
4027 * task to be woken on this_cpu.
4029 if (this_load > 0) {
4030 s64 this_eff_load, prev_eff_load;
4032 this_eff_load = 100;
4033 this_eff_load *= power_of(prev_cpu);
4034 this_eff_load *= this_load +
4035 effective_load(tg, this_cpu, weight, weight);
4037 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
4038 prev_eff_load *= power_of(this_cpu);
4039 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
4041 balanced = this_eff_load <= prev_eff_load;
4046 * If the currently running task will sleep within
4047 * a reasonable amount of time then attract this newly
4050 if (sync && balanced)
4053 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
4054 tl_per_task = cpu_avg_load_per_task(this_cpu);
4057 (this_load <= load &&
4058 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
4060 * This domain has SD_WAKE_AFFINE and
4061 * p is cache cold in this domain, and
4062 * there is no bad imbalance.
4064 schedstat_inc(sd, ttwu_move_affine);
4065 schedstat_inc(p, se.statistics.nr_wakeups_affine);
4073 * find_idlest_group finds and returns the least busy CPU group within the
4076 static struct sched_group *
4077 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
4078 int this_cpu, int load_idx)
4080 struct sched_group *idlest = NULL, *group = sd->groups;
4081 unsigned long min_load = ULONG_MAX, this_load = 0;
4082 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4085 unsigned long load, avg_load;
4089 /* Skip over this group if it has no CPUs allowed */
4090 if (!cpumask_intersects(sched_group_cpus(group),
4091 tsk_cpus_allowed(p)))
4094 local_group = cpumask_test_cpu(this_cpu,
4095 sched_group_cpus(group));
4097 /* Tally up the load of all CPUs in the group */
4100 for_each_cpu(i, sched_group_cpus(group)) {
4101 /* Bias balancing toward cpus of our domain */
4103 load = source_load(i, load_idx);
4105 load = target_load(i, load_idx);
4110 /* Adjust by relative CPU power of the group */
4111 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4114 this_load = avg_load;
4115 } else if (avg_load < min_load) {
4116 min_load = avg_load;
4119 } while (group = group->next, group != sd->groups);
4121 if (!idlest || 100*this_load < imbalance*min_load)
4127 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4130 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4132 unsigned long load, min_load = ULONG_MAX;
4136 /* Traverse only the allowed CPUs */
4137 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4138 load = weighted_cpuload(i);
4140 if (load < min_load || (load == min_load && i == this_cpu)) {
4150 * Try and locate an idle CPU in the sched_domain.
4152 static int select_idle_sibling(struct task_struct *p, int target)
4154 struct sched_domain *sd;
4155 struct sched_group *sg;
4156 int i = task_cpu(p);
4158 if (idle_cpu(target))
4162 * If the prevous cpu is cache affine and idle, don't be stupid.
4164 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4168 * Otherwise, iterate the domains and find an elegible idle cpu.
4170 sd = rcu_dereference(per_cpu(sd_llc, target));
4171 for_each_lower_domain(sd) {
4174 if (!cpumask_intersects(sched_group_cpus(sg),
4175 tsk_cpus_allowed(p)))
4178 for_each_cpu(i, sched_group_cpus(sg)) {
4179 if (i == target || !idle_cpu(i))
4183 target = cpumask_first_and(sched_group_cpus(sg),
4184 tsk_cpus_allowed(p));
4188 } while (sg != sd->groups);
4195 * sched_balance_self: balance the current task (running on cpu) in domains
4196 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4199 * Balance, ie. select the least loaded group.
4201 * Returns the target CPU number, or the same CPU if no balancing is needed.
4203 * preempt must be disabled.
4206 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4208 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4209 int cpu = smp_processor_id();
4211 int want_affine = 0;
4212 int sync = wake_flags & WF_SYNC;
4214 if (p->nr_cpus_allowed == 1)
4217 if (sd_flag & SD_BALANCE_WAKE) {
4218 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4224 for_each_domain(cpu, tmp) {
4225 if (!(tmp->flags & SD_LOAD_BALANCE))
4229 * If both cpu and prev_cpu are part of this domain,
4230 * cpu is a valid SD_WAKE_AFFINE target.
4232 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4233 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4238 if (tmp->flags & sd_flag)
4243 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4246 new_cpu = select_idle_sibling(p, prev_cpu);
4251 int load_idx = sd->forkexec_idx;
4252 struct sched_group *group;
4255 if (!(sd->flags & sd_flag)) {
4260 if (sd_flag & SD_BALANCE_WAKE)
4261 load_idx = sd->wake_idx;
4263 group = find_idlest_group(sd, p, cpu, load_idx);
4269 new_cpu = find_idlest_cpu(group, p, cpu);
4270 if (new_cpu == -1 || new_cpu == cpu) {
4271 /* Now try balancing at a lower domain level of cpu */
4276 /* Now try balancing at a lower domain level of new_cpu */
4278 weight = sd->span_weight;
4280 for_each_domain(cpu, tmp) {
4281 if (weight <= tmp->span_weight)
4283 if (tmp->flags & sd_flag)
4286 /* while loop will break here if sd == NULL */
4295 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4296 * cfs_rq_of(p) references at time of call are still valid and identify the
4297 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4298 * other assumptions, including the state of rq->lock, should be made.
4301 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4303 struct sched_entity *se = &p->se;
4304 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4307 * Load tracking: accumulate removed load so that it can be processed
4308 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4309 * to blocked load iff they have a positive decay-count. It can never
4310 * be negative here since on-rq tasks have decay-count == 0.
4312 if (se->avg.decay_count) {
4313 se->avg.decay_count = -__synchronize_entity_decay(se);
4314 atomic_long_add(se->avg.load_avg_contrib,
4315 &cfs_rq->removed_load);
4318 #endif /* CONFIG_SMP */
4320 static unsigned long
4321 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4323 unsigned long gran = sysctl_sched_wakeup_granularity;
4326 * Since its curr running now, convert the gran from real-time
4327 * to virtual-time in his units.
4329 * By using 'se' instead of 'curr' we penalize light tasks, so
4330 * they get preempted easier. That is, if 'se' < 'curr' then
4331 * the resulting gran will be larger, therefore penalizing the
4332 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4333 * be smaller, again penalizing the lighter task.
4335 * This is especially important for buddies when the leftmost
4336 * task is higher priority than the buddy.
4338 return calc_delta_fair(gran, se);
4342 * Should 'se' preempt 'curr'.
4356 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4358 s64 gran, vdiff = curr->vruntime - se->vruntime;
4363 gran = wakeup_gran(curr, se);
4370 static void set_last_buddy(struct sched_entity *se)
4372 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4375 for_each_sched_entity(se)
4376 cfs_rq_of(se)->last = se;
4379 static void set_next_buddy(struct sched_entity *se)
4381 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4384 for_each_sched_entity(se)
4385 cfs_rq_of(se)->next = se;
4388 static void set_skip_buddy(struct sched_entity *se)
4390 for_each_sched_entity(se)
4391 cfs_rq_of(se)->skip = se;
4395 * Preempt the current task with a newly woken task if needed:
4397 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4399 struct task_struct *curr = rq->curr;
4400 struct sched_entity *se = &curr->se, *pse = &p->se;
4401 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4402 int scale = cfs_rq->nr_running >= sched_nr_latency;
4403 int next_buddy_marked = 0;
4405 if (unlikely(se == pse))
4409 * This is possible from callers such as move_task(), in which we
4410 * unconditionally check_prempt_curr() after an enqueue (which may have
4411 * lead to a throttle). This both saves work and prevents false
4412 * next-buddy nomination below.
4414 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4417 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4418 set_next_buddy(pse);
4419 next_buddy_marked = 1;
4423 * We can come here with TIF_NEED_RESCHED already set from new task
4426 * Note: this also catches the edge-case of curr being in a throttled
4427 * group (e.g. via set_curr_task), since update_curr() (in the
4428 * enqueue of curr) will have resulted in resched being set. This
4429 * prevents us from potentially nominating it as a false LAST_BUDDY
4432 if (test_tsk_need_resched(curr))
4435 /* Idle tasks are by definition preempted by non-idle tasks. */
4436 if (unlikely(curr->policy == SCHED_IDLE) &&
4437 likely(p->policy != SCHED_IDLE))
4441 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4442 * is driven by the tick):
4444 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4447 find_matching_se(&se, &pse);
4448 update_curr(cfs_rq_of(se));
4450 if (wakeup_preempt_entity(se, pse) == 1) {
4452 * Bias pick_next to pick the sched entity that is
4453 * triggering this preemption.
4455 if (!next_buddy_marked)
4456 set_next_buddy(pse);
4465 * Only set the backward buddy when the current task is still
4466 * on the rq. This can happen when a wakeup gets interleaved
4467 * with schedule on the ->pre_schedule() or idle_balance()
4468 * point, either of which can * drop the rq lock.
4470 * Also, during early boot the idle thread is in the fair class,
4471 * for obvious reasons its a bad idea to schedule back to it.
4473 if (unlikely(!se->on_rq || curr == rq->idle))
4476 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4480 static struct task_struct *pick_next_task_fair(struct rq *rq)
4482 struct task_struct *p;
4483 struct cfs_rq *cfs_rq = &rq->cfs;
4484 struct sched_entity *se;
4486 if (!cfs_rq->nr_running)
4490 se = pick_next_entity(cfs_rq);
4491 set_next_entity(cfs_rq, se);
4492 cfs_rq = group_cfs_rq(se);
4496 if (hrtick_enabled(rq))
4497 hrtick_start_fair(rq, p);
4503 * Account for a descheduled task:
4505 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4507 struct sched_entity *se = &prev->se;
4508 struct cfs_rq *cfs_rq;
4510 for_each_sched_entity(se) {
4511 cfs_rq = cfs_rq_of(se);
4512 put_prev_entity(cfs_rq, se);
4517 * sched_yield() is very simple
4519 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4521 static void yield_task_fair(struct rq *rq)
4523 struct task_struct *curr = rq->curr;
4524 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4525 struct sched_entity *se = &curr->se;
4528 * Are we the only task in the tree?
4530 if (unlikely(rq->nr_running == 1))
4533 clear_buddies(cfs_rq, se);
4535 if (curr->policy != SCHED_BATCH) {
4536 update_rq_clock(rq);
4538 * Update run-time statistics of the 'current'.
4540 update_curr(cfs_rq);
4542 * Tell update_rq_clock() that we've just updated,
4543 * so we don't do microscopic update in schedule()
4544 * and double the fastpath cost.
4546 rq->skip_clock_update = 1;
4552 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4554 struct sched_entity *se = &p->se;
4556 /* throttled hierarchies are not runnable */
4557 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4560 /* Tell the scheduler that we'd really like pse to run next. */
4563 yield_task_fair(rq);
4569 /**************************************************
4570 * Fair scheduling class load-balancing methods.
4574 * The purpose of load-balancing is to achieve the same basic fairness the
4575 * per-cpu scheduler provides, namely provide a proportional amount of compute
4576 * time to each task. This is expressed in the following equation:
4578 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4580 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4581 * W_i,0 is defined as:
4583 * W_i,0 = \Sum_j w_i,j (2)
4585 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4586 * is derived from the nice value as per prio_to_weight[].
4588 * The weight average is an exponential decay average of the instantaneous
4591 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4593 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4594 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4595 * can also include other factors [XXX].
4597 * To achieve this balance we define a measure of imbalance which follows
4598 * directly from (1):
4600 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4602 * We them move tasks around to minimize the imbalance. In the continuous
4603 * function space it is obvious this converges, in the discrete case we get
4604 * a few fun cases generally called infeasible weight scenarios.
4607 * - infeasible weights;
4608 * - local vs global optima in the discrete case. ]
4613 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4614 * for all i,j solution, we create a tree of cpus that follows the hardware
4615 * topology where each level pairs two lower groups (or better). This results
4616 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4617 * tree to only the first of the previous level and we decrease the frequency
4618 * of load-balance at each level inv. proportional to the number of cpus in
4624 * \Sum { --- * --- * 2^i } = O(n) (5)
4626 * `- size of each group
4627 * | | `- number of cpus doing load-balance
4629 * `- sum over all levels
4631 * Coupled with a limit on how many tasks we can migrate every balance pass,
4632 * this makes (5) the runtime complexity of the balancer.
4634 * An important property here is that each CPU is still (indirectly) connected
4635 * to every other cpu in at most O(log n) steps:
4637 * The adjacency matrix of the resulting graph is given by:
4640 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4643 * And you'll find that:
4645 * A^(log_2 n)_i,j != 0 for all i,j (7)
4647 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4648 * The task movement gives a factor of O(m), giving a convergence complexity
4651 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4656 * In order to avoid CPUs going idle while there's still work to do, new idle
4657 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4658 * tree itself instead of relying on other CPUs to bring it work.
4660 * This adds some complexity to both (5) and (8) but it reduces the total idle
4668 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4671 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4676 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4678 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4680 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4683 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4684 * rewrite all of this once again.]
4687 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4689 enum fbq_type { regular, remote, all };
4691 #define LBF_ALL_PINNED 0x01
4692 #define LBF_NEED_BREAK 0x02
4693 #define LBF_DST_PINNED 0x04
4694 #define LBF_SOME_PINNED 0x08
4697 struct sched_domain *sd;
4705 struct cpumask *dst_grpmask;
4707 enum cpu_idle_type idle;
4709 /* The set of CPUs under consideration for load-balancing */
4710 struct cpumask *cpus;
4715 unsigned int loop_break;
4716 unsigned int loop_max;
4718 enum fbq_type fbq_type;
4722 * move_task - move a task from one runqueue to another runqueue.
4723 * Both runqueues must be locked.
4725 static void move_task(struct task_struct *p, struct lb_env *env)
4727 deactivate_task(env->src_rq, p, 0);
4728 set_task_cpu(p, env->dst_cpu);
4729 activate_task(env->dst_rq, p, 0);
4730 check_preempt_curr(env->dst_rq, p, 0);
4734 * Is this task likely cache-hot:
4737 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4741 if (p->sched_class != &fair_sched_class)
4744 if (unlikely(p->policy == SCHED_IDLE))
4748 * Buddy candidates are cache hot:
4750 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4751 (&p->se == cfs_rq_of(&p->se)->next ||
4752 &p->se == cfs_rq_of(&p->se)->last))
4755 if (sysctl_sched_migration_cost == -1)
4757 if (sysctl_sched_migration_cost == 0)
4760 delta = now - p->se.exec_start;
4762 return delta < (s64)sysctl_sched_migration_cost;
4765 #ifdef CONFIG_NUMA_BALANCING
4766 /* Returns true if the destination node has incurred more faults */
4767 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4769 int src_nid, dst_nid;
4771 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4772 !(env->sd->flags & SD_NUMA)) {
4776 src_nid = cpu_to_node(env->src_cpu);
4777 dst_nid = cpu_to_node(env->dst_cpu);
4779 if (src_nid == dst_nid)
4782 /* Always encourage migration to the preferred node. */
4783 if (dst_nid == p->numa_preferred_nid)
4786 /* If both task and group weight improve, this move is a winner. */
4787 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4788 group_weight(p, dst_nid) > group_weight(p, src_nid))
4795 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4797 int src_nid, dst_nid;
4799 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4802 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4805 src_nid = cpu_to_node(env->src_cpu);
4806 dst_nid = cpu_to_node(env->dst_cpu);
4808 if (src_nid == dst_nid)
4811 /* Migrating away from the preferred node is always bad. */
4812 if (src_nid == p->numa_preferred_nid)
4815 /* If either task or group weight get worse, don't do it. */
4816 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4817 group_weight(p, dst_nid) < group_weight(p, src_nid))
4824 static inline bool migrate_improves_locality(struct task_struct *p,
4830 static inline bool migrate_degrades_locality(struct task_struct *p,
4838 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4841 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4843 int tsk_cache_hot = 0;
4845 * We do not migrate tasks that are:
4846 * 1) throttled_lb_pair, or
4847 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4848 * 3) running (obviously), or
4849 * 4) are cache-hot on their current CPU.
4851 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4854 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4857 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4859 env->flags |= LBF_SOME_PINNED;
4862 * Remember if this task can be migrated to any other cpu in
4863 * our sched_group. We may want to revisit it if we couldn't
4864 * meet load balance goals by pulling other tasks on src_cpu.
4866 * Also avoid computing new_dst_cpu if we have already computed
4867 * one in current iteration.
4869 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4872 /* Prevent to re-select dst_cpu via env's cpus */
4873 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4874 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4875 env->flags |= LBF_DST_PINNED;
4876 env->new_dst_cpu = cpu;
4884 /* Record that we found atleast one task that could run on dst_cpu */
4885 env->flags &= ~LBF_ALL_PINNED;
4887 if (task_running(env->src_rq, p)) {
4888 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4893 * Aggressive migration if:
4894 * 1) destination numa is preferred
4895 * 2) task is cache cold, or
4896 * 3) too many balance attempts have failed.
4898 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4900 tsk_cache_hot = migrate_degrades_locality(p, env);
4902 if (migrate_improves_locality(p, env)) {
4903 #ifdef CONFIG_SCHEDSTATS
4904 if (tsk_cache_hot) {
4905 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4906 schedstat_inc(p, se.statistics.nr_forced_migrations);
4912 if (!tsk_cache_hot ||
4913 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4915 if (tsk_cache_hot) {
4916 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4917 schedstat_inc(p, se.statistics.nr_forced_migrations);
4923 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4928 * move_one_task tries to move exactly one task from busiest to this_rq, as
4929 * part of active balancing operations within "domain".
4930 * Returns 1 if successful and 0 otherwise.
4932 * Called with both runqueues locked.
4934 static int move_one_task(struct lb_env *env)
4936 struct task_struct *p, *n;
4938 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4939 if (!can_migrate_task(p, env))
4944 * Right now, this is only the second place move_task()
4945 * is called, so we can safely collect move_task()
4946 * stats here rather than inside move_task().
4948 schedstat_inc(env->sd, lb_gained[env->idle]);
4954 static const unsigned int sched_nr_migrate_break = 32;
4957 * move_tasks tries to move up to imbalance weighted load from busiest to
4958 * this_rq, as part of a balancing operation within domain "sd".
4959 * Returns 1 if successful and 0 otherwise.
4961 * Called with both runqueues locked.
4963 static int move_tasks(struct lb_env *env)
4965 struct list_head *tasks = &env->src_rq->cfs_tasks;
4966 struct task_struct *p;
4970 if (env->imbalance <= 0)
4973 while (!list_empty(tasks)) {
4974 p = list_first_entry(tasks, struct task_struct, se.group_node);
4977 /* We've more or less seen every task there is, call it quits */
4978 if (env->loop > env->loop_max)
4981 /* take a breather every nr_migrate tasks */
4982 if (env->loop > env->loop_break) {
4983 env->loop_break += sched_nr_migrate_break;
4984 env->flags |= LBF_NEED_BREAK;
4988 if (!can_migrate_task(p, env))
4991 load = task_h_load(p);
4993 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4996 if ((load / 2) > env->imbalance)
5001 env->imbalance -= load;
5003 #ifdef CONFIG_PREEMPT
5005 * NEWIDLE balancing is a source of latency, so preemptible
5006 * kernels will stop after the first task is pulled to minimize
5007 * the critical section.
5009 if (env->idle == CPU_NEWLY_IDLE)
5014 * We only want to steal up to the prescribed amount of
5017 if (env->imbalance <= 0)
5022 list_move_tail(&p->se.group_node, tasks);
5026 * Right now, this is one of only two places move_task() is called,
5027 * so we can safely collect move_task() stats here rather than
5028 * inside move_task().
5030 schedstat_add(env->sd, lb_gained[env->idle], pulled);
5035 #ifdef CONFIG_FAIR_GROUP_SCHED
5037 * update tg->load_weight by folding this cpu's load_avg
5039 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
5041 struct sched_entity *se = tg->se[cpu];
5042 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
5044 /* throttled entities do not contribute to load */
5045 if (throttled_hierarchy(cfs_rq))
5048 update_cfs_rq_blocked_load(cfs_rq, 1);
5051 update_entity_load_avg(se, 1);
5053 * We pivot on our runnable average having decayed to zero for
5054 * list removal. This generally implies that all our children
5055 * have also been removed (modulo rounding error or bandwidth
5056 * control); however, such cases are rare and we can fix these
5059 * TODO: fix up out-of-order children on enqueue.
5061 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
5062 list_del_leaf_cfs_rq(cfs_rq);
5064 struct rq *rq = rq_of(cfs_rq);
5065 update_rq_runnable_avg(rq, rq->nr_running);
5069 static void update_blocked_averages(int cpu)
5071 struct rq *rq = cpu_rq(cpu);
5072 struct cfs_rq *cfs_rq;
5073 unsigned long flags;
5075 raw_spin_lock_irqsave(&rq->lock, flags);
5076 update_rq_clock(rq);
5078 * Iterates the task_group tree in a bottom up fashion, see
5079 * list_add_leaf_cfs_rq() for details.
5081 for_each_leaf_cfs_rq(rq, cfs_rq) {
5083 * Note: We may want to consider periodically releasing
5084 * rq->lock about these updates so that creating many task
5085 * groups does not result in continually extending hold time.
5087 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5090 raw_spin_unlock_irqrestore(&rq->lock, flags);
5094 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5095 * This needs to be done in a top-down fashion because the load of a child
5096 * group is a fraction of its parents load.
5098 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5100 struct rq *rq = rq_of(cfs_rq);
5101 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5102 unsigned long now = jiffies;
5105 if (cfs_rq->last_h_load_update == now)
5108 cfs_rq->h_load_next = NULL;
5109 for_each_sched_entity(se) {
5110 cfs_rq = cfs_rq_of(se);
5111 cfs_rq->h_load_next = se;
5112 if (cfs_rq->last_h_load_update == now)
5117 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5118 cfs_rq->last_h_load_update = now;
5121 while ((se = cfs_rq->h_load_next) != NULL) {
5122 load = cfs_rq->h_load;
5123 load = div64_ul(load * se->avg.load_avg_contrib,
5124 cfs_rq->runnable_load_avg + 1);
5125 cfs_rq = group_cfs_rq(se);
5126 cfs_rq->h_load = load;
5127 cfs_rq->last_h_load_update = now;
5131 static unsigned long task_h_load(struct task_struct *p)
5133 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5135 update_cfs_rq_h_load(cfs_rq);
5136 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5137 cfs_rq->runnable_load_avg + 1);
5140 static inline void update_blocked_averages(int cpu)
5144 static unsigned long task_h_load(struct task_struct *p)
5146 return p->se.avg.load_avg_contrib;
5150 /********** Helpers for find_busiest_group ************************/
5152 * sg_lb_stats - stats of a sched_group required for load_balancing
5154 struct sg_lb_stats {
5155 unsigned long avg_load; /*Avg load across the CPUs of the group */
5156 unsigned long group_load; /* Total load over the CPUs of the group */
5157 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5158 unsigned long load_per_task;
5159 unsigned long group_power;
5160 unsigned int sum_nr_running; /* Nr tasks running in the group */
5161 unsigned int group_capacity;
5162 unsigned int idle_cpus;
5163 unsigned int group_weight;
5164 int group_imb; /* Is there an imbalance in the group ? */
5165 int group_has_capacity; /* Is there extra capacity in the group? */
5166 #ifdef CONFIG_NUMA_BALANCING
5167 unsigned int nr_numa_running;
5168 unsigned int nr_preferred_running;
5173 * sd_lb_stats - Structure to store the statistics of a sched_domain
5174 * during load balancing.
5176 struct sd_lb_stats {
5177 struct sched_group *busiest; /* Busiest group in this sd */
5178 struct sched_group *local; /* Local group in this sd */
5179 unsigned long total_load; /* Total load of all groups in sd */
5180 unsigned long total_pwr; /* Total power of all groups in sd */
5181 unsigned long avg_load; /* Average load across all groups in sd */
5183 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5184 struct sg_lb_stats local_stat; /* Statistics of the local group */
5187 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5190 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5191 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5192 * We must however clear busiest_stat::avg_load because
5193 * update_sd_pick_busiest() reads this before assignment.
5195 *sds = (struct sd_lb_stats){
5207 * get_sd_load_idx - Obtain the load index for a given sched domain.
5208 * @sd: The sched_domain whose load_idx is to be obtained.
5209 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
5211 * Return: The load index.
5213 static inline int get_sd_load_idx(struct sched_domain *sd,
5214 enum cpu_idle_type idle)
5220 load_idx = sd->busy_idx;
5223 case CPU_NEWLY_IDLE:
5224 load_idx = sd->newidle_idx;
5227 load_idx = sd->idle_idx;
5234 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5236 return SCHED_POWER_SCALE;
5239 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5241 return default_scale_freq_power(sd, cpu);
5244 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5246 unsigned long weight = sd->span_weight;
5247 unsigned long smt_gain = sd->smt_gain;
5254 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5256 return default_scale_smt_power(sd, cpu);
5259 static unsigned long scale_rt_power(int cpu)
5261 struct rq *rq = cpu_rq(cpu);
5262 u64 total, available, age_stamp, avg;
5265 * Since we're reading these variables without serialization make sure
5266 * we read them once before doing sanity checks on them.
5268 age_stamp = ACCESS_ONCE(rq->age_stamp);
5269 avg = ACCESS_ONCE(rq->rt_avg);
5271 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5273 if (unlikely(total < avg)) {
5274 /* Ensures that power won't end up being negative */
5277 available = total - avg;
5280 if (unlikely((s64)total < SCHED_POWER_SCALE))
5281 total = SCHED_POWER_SCALE;
5283 total >>= SCHED_POWER_SHIFT;
5285 return div_u64(available, total);
5288 static void update_cpu_power(struct sched_domain *sd, int cpu)
5290 unsigned long weight = sd->span_weight;
5291 unsigned long power = SCHED_POWER_SCALE;
5292 struct sched_group *sdg = sd->groups;
5294 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5295 if (sched_feat(ARCH_POWER))
5296 power *= arch_scale_smt_power(sd, cpu);
5298 power *= default_scale_smt_power(sd, cpu);
5300 power >>= SCHED_POWER_SHIFT;
5303 sdg->sgp->power_orig = power;
5305 if (sched_feat(ARCH_POWER))
5306 power *= arch_scale_freq_power(sd, cpu);
5308 power *= default_scale_freq_power(sd, cpu);
5310 power >>= SCHED_POWER_SHIFT;
5312 power *= scale_rt_power(cpu);
5313 power >>= SCHED_POWER_SHIFT;
5318 cpu_rq(cpu)->cpu_power = power;
5319 sdg->sgp->power = power;
5322 void update_group_power(struct sched_domain *sd, int cpu)
5324 struct sched_domain *child = sd->child;
5325 struct sched_group *group, *sdg = sd->groups;
5326 unsigned long power, power_orig;
5327 unsigned long interval;
5329 interval = msecs_to_jiffies(sd->balance_interval);
5330 interval = clamp(interval, 1UL, max_load_balance_interval);
5331 sdg->sgp->next_update = jiffies + interval;
5334 update_cpu_power(sd, cpu);
5338 power_orig = power = 0;
5340 if (child->flags & SD_OVERLAP) {
5342 * SD_OVERLAP domains cannot assume that child groups
5343 * span the current group.
5346 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5347 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5349 power_orig += sg->sgp->power_orig;
5350 power += sg->sgp->power;
5354 * !SD_OVERLAP domains can assume that child groups
5355 * span the current group.
5358 group = child->groups;
5360 power_orig += group->sgp->power_orig;
5361 power += group->sgp->power;
5362 group = group->next;
5363 } while (group != child->groups);
5366 sdg->sgp->power_orig = power_orig;
5367 sdg->sgp->power = power;
5371 * Try and fix up capacity for tiny siblings, this is needed when
5372 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5373 * which on its own isn't powerful enough.
5375 * See update_sd_pick_busiest() and check_asym_packing().
5378 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5381 * Only siblings can have significantly less than SCHED_POWER_SCALE
5383 if (!(sd->flags & SD_SHARE_CPUPOWER))
5387 * If ~90% of the cpu_power is still there, we're good.
5389 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5396 * Group imbalance indicates (and tries to solve) the problem where balancing
5397 * groups is inadequate due to tsk_cpus_allowed() constraints.
5399 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5400 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5403 * { 0 1 2 3 } { 4 5 6 7 }
5406 * If we were to balance group-wise we'd place two tasks in the first group and
5407 * two tasks in the second group. Clearly this is undesired as it will overload
5408 * cpu 3 and leave one of the cpus in the second group unused.
5410 * The current solution to this issue is detecting the skew in the first group
5411 * by noticing the lower domain failed to reach balance and had difficulty
5412 * moving tasks due to affinity constraints.
5414 * When this is so detected; this group becomes a candidate for busiest; see
5415 * update_sd_pick_busiest(). And calculate_imbalance() and
5416 * find_busiest_group() avoid some of the usual balance conditions to allow it
5417 * to create an effective group imbalance.
5419 * This is a somewhat tricky proposition since the next run might not find the
5420 * group imbalance and decide the groups need to be balanced again. A most
5421 * subtle and fragile situation.
5424 static inline int sg_imbalanced(struct sched_group *group)
5426 return group->sgp->imbalance;
5430 * Compute the group capacity.
5432 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5433 * first dividing out the smt factor and computing the actual number of cores
5434 * and limit power unit capacity with that.
5436 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5438 unsigned int capacity, smt, cpus;
5439 unsigned int power, power_orig;
5441 power = group->sgp->power;
5442 power_orig = group->sgp->power_orig;
5443 cpus = group->group_weight;
5445 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5446 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5447 capacity = cpus / smt; /* cores */
5449 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5451 capacity = fix_small_capacity(env->sd, group);
5457 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5458 * @env: The load balancing environment.
5459 * @group: sched_group whose statistics are to be updated.
5460 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5461 * @local_group: Does group contain this_cpu.
5462 * @sgs: variable to hold the statistics for this group.
5464 static inline void update_sg_lb_stats(struct lb_env *env,
5465 struct sched_group *group, int load_idx,
5466 int local_group, struct sg_lb_stats *sgs)
5468 unsigned long nr_running;
5472 memset(sgs, 0, sizeof(*sgs));
5474 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5475 struct rq *rq = cpu_rq(i);
5477 nr_running = rq->nr_running;
5479 /* Bias balancing toward cpus of our domain */
5481 load = target_load(i, load_idx);
5483 load = source_load(i, load_idx);
5485 sgs->group_load += load;
5486 sgs->sum_nr_running += nr_running;
5487 #ifdef CONFIG_NUMA_BALANCING
5488 sgs->nr_numa_running += rq->nr_numa_running;
5489 sgs->nr_preferred_running += rq->nr_preferred_running;
5491 sgs->sum_weighted_load += weighted_cpuload(i);
5496 /* Adjust by relative CPU power of the group */
5497 sgs->group_power = group->sgp->power;
5498 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5500 if (sgs->sum_nr_running)
5501 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5503 sgs->group_weight = group->group_weight;
5505 sgs->group_imb = sg_imbalanced(group);
5506 sgs->group_capacity = sg_capacity(env, group);
5508 if (sgs->group_capacity > sgs->sum_nr_running)
5509 sgs->group_has_capacity = 1;
5513 * update_sd_pick_busiest - return 1 on busiest group
5514 * @env: The load balancing environment.
5515 * @sds: sched_domain statistics
5516 * @sg: sched_group candidate to be checked for being the busiest
5517 * @sgs: sched_group statistics
5519 * Determine if @sg is a busier group than the previously selected
5522 * Return: %true if @sg is a busier group than the previously selected
5523 * busiest group. %false otherwise.
5525 static bool update_sd_pick_busiest(struct lb_env *env,
5526 struct sd_lb_stats *sds,
5527 struct sched_group *sg,
5528 struct sg_lb_stats *sgs)
5530 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5533 if (sgs->sum_nr_running > sgs->group_capacity)
5540 * ASYM_PACKING needs to move all the work to the lowest
5541 * numbered CPUs in the group, therefore mark all groups
5542 * higher than ourself as busy.
5544 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5545 env->dst_cpu < group_first_cpu(sg)) {
5549 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5556 #ifdef CONFIG_NUMA_BALANCING
5557 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5559 if (sgs->sum_nr_running > sgs->nr_numa_running)
5561 if (sgs->sum_nr_running > sgs->nr_preferred_running)
5566 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5568 if (rq->nr_running > rq->nr_numa_running)
5570 if (rq->nr_running > rq->nr_preferred_running)
5575 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
5580 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
5584 #endif /* CONFIG_NUMA_BALANCING */
5587 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5588 * @env: The load balancing environment.
5589 * @sds: variable to hold the statistics for this sched_domain.
5591 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
5593 struct sched_domain *child = env->sd->child;
5594 struct sched_group *sg = env->sd->groups;
5595 struct sg_lb_stats tmp_sgs;
5596 int load_idx, prefer_sibling = 0;
5598 if (child && child->flags & SD_PREFER_SIBLING)
5601 load_idx = get_sd_load_idx(env->sd, env->idle);
5604 struct sg_lb_stats *sgs = &tmp_sgs;
5607 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5610 sgs = &sds->local_stat;
5612 if (env->idle != CPU_NEWLY_IDLE ||
5613 time_after_eq(jiffies, sg->sgp->next_update))
5614 update_group_power(env->sd, env->dst_cpu);
5617 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5623 * In case the child domain prefers tasks go to siblings
5624 * first, lower the sg capacity to one so that we'll try
5625 * and move all the excess tasks away. We lower the capacity
5626 * of a group only if the local group has the capacity to fit
5627 * these excess tasks, i.e. nr_running < group_capacity. The
5628 * extra check prevents the case where you always pull from the
5629 * heaviest group when it is already under-utilized (possible
5630 * with a large weight task outweighs the tasks on the system).
5632 if (prefer_sibling && sds->local &&
5633 sds->local_stat.group_has_capacity)
5634 sgs->group_capacity = min(sgs->group_capacity, 1U);
5636 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5638 sds->busiest_stat = *sgs;
5642 /* Now, start updating sd_lb_stats */
5643 sds->total_load += sgs->group_load;
5644 sds->total_pwr += sgs->group_power;
5647 } while (sg != env->sd->groups);
5649 if (env->sd->flags & SD_NUMA)
5650 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
5654 * check_asym_packing - Check to see if the group is packed into the
5657 * This is primarily intended to used at the sibling level. Some
5658 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5659 * case of POWER7, it can move to lower SMT modes only when higher
5660 * threads are idle. When in lower SMT modes, the threads will
5661 * perform better since they share less core resources. Hence when we
5662 * have idle threads, we want them to be the higher ones.
5664 * This packing function is run on idle threads. It checks to see if
5665 * the busiest CPU in this domain (core in the P7 case) has a higher
5666 * CPU number than the packing function is being run on. Here we are
5667 * assuming lower CPU number will be equivalent to lower a SMT thread
5670 * Return: 1 when packing is required and a task should be moved to
5671 * this CPU. The amount of the imbalance is returned in *imbalance.
5673 * @env: The load balancing environment.
5674 * @sds: Statistics of the sched_domain which is to be packed
5676 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5680 if (!(env->sd->flags & SD_ASYM_PACKING))
5686 busiest_cpu = group_first_cpu(sds->busiest);
5687 if (env->dst_cpu > busiest_cpu)
5690 env->imbalance = DIV_ROUND_CLOSEST(
5691 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5698 * fix_small_imbalance - Calculate the minor imbalance that exists
5699 * amongst the groups of a sched_domain, during
5701 * @env: The load balancing environment.
5702 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5705 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5707 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5708 unsigned int imbn = 2;
5709 unsigned long scaled_busy_load_per_task;
5710 struct sg_lb_stats *local, *busiest;
5712 local = &sds->local_stat;
5713 busiest = &sds->busiest_stat;
5715 if (!local->sum_nr_running)
5716 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5717 else if (busiest->load_per_task > local->load_per_task)
5720 scaled_busy_load_per_task =
5721 (busiest->load_per_task * SCHED_POWER_SCALE) /
5722 busiest->group_power;
5724 if (busiest->avg_load + scaled_busy_load_per_task >=
5725 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5726 env->imbalance = busiest->load_per_task;
5731 * OK, we don't have enough imbalance to justify moving tasks,
5732 * however we may be able to increase total CPU power used by
5736 pwr_now += busiest->group_power *
5737 min(busiest->load_per_task, busiest->avg_load);
5738 pwr_now += local->group_power *
5739 min(local->load_per_task, local->avg_load);
5740 pwr_now /= SCHED_POWER_SCALE;
5742 /* Amount of load we'd subtract */
5743 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5744 busiest->group_power;
5745 if (busiest->avg_load > tmp) {
5746 pwr_move += busiest->group_power *
5747 min(busiest->load_per_task,
5748 busiest->avg_load - tmp);
5751 /* Amount of load we'd add */
5752 if (busiest->avg_load * busiest->group_power <
5753 busiest->load_per_task * SCHED_POWER_SCALE) {
5754 tmp = (busiest->avg_load * busiest->group_power) /
5757 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5760 pwr_move += local->group_power *
5761 min(local->load_per_task, local->avg_load + tmp);
5762 pwr_move /= SCHED_POWER_SCALE;
5764 /* Move if we gain throughput */
5765 if (pwr_move > pwr_now)
5766 env->imbalance = busiest->load_per_task;
5770 * calculate_imbalance - Calculate the amount of imbalance present within the
5771 * groups of a given sched_domain during load balance.
5772 * @env: load balance environment
5773 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5775 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5777 unsigned long max_pull, load_above_capacity = ~0UL;
5778 struct sg_lb_stats *local, *busiest;
5780 local = &sds->local_stat;
5781 busiest = &sds->busiest_stat;
5783 if (busiest->group_imb) {
5785 * In the group_imb case we cannot rely on group-wide averages
5786 * to ensure cpu-load equilibrium, look at wider averages. XXX
5788 busiest->load_per_task =
5789 min(busiest->load_per_task, sds->avg_load);
5793 * In the presence of smp nice balancing, certain scenarios can have
5794 * max load less than avg load(as we skip the groups at or below
5795 * its cpu_power, while calculating max_load..)
5797 if (busiest->avg_load <= sds->avg_load ||
5798 local->avg_load >= sds->avg_load) {
5800 return fix_small_imbalance(env, sds);
5803 if (!busiest->group_imb) {
5805 * Don't want to pull so many tasks that a group would go idle.
5806 * Except of course for the group_imb case, since then we might
5807 * have to drop below capacity to reach cpu-load equilibrium.
5809 load_above_capacity =
5810 (busiest->sum_nr_running - busiest->group_capacity);
5812 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5813 load_above_capacity /= busiest->group_power;
5817 * We're trying to get all the cpus to the average_load, so we don't
5818 * want to push ourselves above the average load, nor do we wish to
5819 * reduce the max loaded cpu below the average load. At the same time,
5820 * we also don't want to reduce the group load below the group capacity
5821 * (so that we can implement power-savings policies etc). Thus we look
5822 * for the minimum possible imbalance.
5824 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5826 /* How much load to actually move to equalise the imbalance */
5827 env->imbalance = min(
5828 max_pull * busiest->group_power,
5829 (sds->avg_load - local->avg_load) * local->group_power
5830 ) / SCHED_POWER_SCALE;
5833 * if *imbalance is less than the average load per runnable task
5834 * there is no guarantee that any tasks will be moved so we'll have
5835 * a think about bumping its value to force at least one task to be
5838 if (env->imbalance < busiest->load_per_task)
5839 return fix_small_imbalance(env, sds);
5842 /******* find_busiest_group() helpers end here *********************/
5845 * find_busiest_group - Returns the busiest group within the sched_domain
5846 * if there is an imbalance. If there isn't an imbalance, and
5847 * the user has opted for power-savings, it returns a group whose
5848 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5849 * such a group exists.
5851 * Also calculates the amount of weighted load which should be moved
5852 * to restore balance.
5854 * @env: The load balancing environment.
5856 * Return: - The busiest group if imbalance exists.
5857 * - If no imbalance and user has opted for power-savings balance,
5858 * return the least loaded group whose CPUs can be
5859 * put to idle by rebalancing its tasks onto our group.
5861 static struct sched_group *find_busiest_group(struct lb_env *env)
5863 struct sg_lb_stats *local, *busiest;
5864 struct sd_lb_stats sds;
5866 init_sd_lb_stats(&sds);
5869 * Compute the various statistics relavent for load balancing at
5872 update_sd_lb_stats(env, &sds);
5873 local = &sds.local_stat;
5874 busiest = &sds.busiest_stat;
5876 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5877 check_asym_packing(env, &sds))
5880 /* There is no busy sibling group to pull tasks from */
5881 if (!sds.busiest || busiest->sum_nr_running == 0)
5884 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5887 * If the busiest group is imbalanced the below checks don't
5888 * work because they assume all things are equal, which typically
5889 * isn't true due to cpus_allowed constraints and the like.
5891 if (busiest->group_imb)
5894 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5895 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5896 !busiest->group_has_capacity)
5900 * If the local group is more busy than the selected busiest group
5901 * don't try and pull any tasks.
5903 if (local->avg_load >= busiest->avg_load)
5907 * Don't pull any tasks if this group is already above the domain
5910 if (local->avg_load >= sds.avg_load)
5913 if (env->idle == CPU_IDLE) {
5915 * This cpu is idle. If the busiest group load doesn't
5916 * have more tasks than the number of available cpu's and
5917 * there is no imbalance between this and busiest group
5918 * wrt to idle cpu's, it is balanced.
5920 if ((local->idle_cpus < busiest->idle_cpus) &&
5921 busiest->sum_nr_running <= busiest->group_weight)
5925 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5926 * imbalance_pct to be conservative.
5928 if (100 * busiest->avg_load <=
5929 env->sd->imbalance_pct * local->avg_load)
5934 /* Looks like there is an imbalance. Compute it */
5935 calculate_imbalance(env, &sds);
5944 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5946 static struct rq *find_busiest_queue(struct lb_env *env,
5947 struct sched_group *group)
5949 struct rq *busiest = NULL, *rq;
5950 unsigned long busiest_load = 0, busiest_power = 1;
5953 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5954 unsigned long power, capacity, wl;
5958 rt = fbq_classify_rq(rq);
5961 * We classify groups/runqueues into three groups:
5962 * - regular: there are !numa tasks
5963 * - remote: there are numa tasks that run on the 'wrong' node
5964 * - all: there is no distinction
5966 * In order to avoid migrating ideally placed numa tasks,
5967 * ignore those when there's better options.
5969 * If we ignore the actual busiest queue to migrate another
5970 * task, the next balance pass can still reduce the busiest
5971 * queue by moving tasks around inside the node.
5973 * If we cannot move enough load due to this classification
5974 * the next pass will adjust the group classification and
5975 * allow migration of more tasks.
5977 * Both cases only affect the total convergence complexity.
5979 if (rt > env->fbq_type)
5982 power = power_of(i);
5983 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
5985 capacity = fix_small_capacity(env->sd, group);
5987 wl = weighted_cpuload(i);
5990 * When comparing with imbalance, use weighted_cpuload()
5991 * which is not scaled with the cpu power.
5993 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5997 * For the load comparisons with the other cpu's, consider
5998 * the weighted_cpuload() scaled with the cpu power, so that
5999 * the load can be moved away from the cpu that is potentially
6000 * running at a lower capacity.
6002 * Thus we're looking for max(wl_i / power_i), crosswise
6003 * multiplication to rid ourselves of the division works out
6004 * to: wl_i * power_j > wl_j * power_i; where j is our
6007 if (wl * busiest_power > busiest_load * power) {
6009 busiest_power = power;
6018 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
6019 * so long as it is large enough.
6021 #define MAX_PINNED_INTERVAL 512
6023 /* Working cpumask for load_balance and load_balance_newidle. */
6024 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6026 static int need_active_balance(struct lb_env *env)
6028 struct sched_domain *sd = env->sd;
6030 if (env->idle == CPU_NEWLY_IDLE) {
6033 * ASYM_PACKING needs to force migrate tasks from busy but
6034 * higher numbered CPUs in order to pack all tasks in the
6035 * lowest numbered CPUs.
6037 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
6041 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
6044 static int active_load_balance_cpu_stop(void *data);
6046 static int should_we_balance(struct lb_env *env)
6048 struct sched_group *sg = env->sd->groups;
6049 struct cpumask *sg_cpus, *sg_mask;
6050 int cpu, balance_cpu = -1;
6053 * In the newly idle case, we will allow all the cpu's
6054 * to do the newly idle load balance.
6056 if (env->idle == CPU_NEWLY_IDLE)
6059 sg_cpus = sched_group_cpus(sg);
6060 sg_mask = sched_group_mask(sg);
6061 /* Try to find first idle cpu */
6062 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
6063 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
6070 if (balance_cpu == -1)
6071 balance_cpu = group_balance_cpu(sg);
6074 * First idle cpu or the first cpu(busiest) in this sched group
6075 * is eligible for doing load balancing at this and above domains.
6077 return balance_cpu == env->dst_cpu;
6081 * Check this_cpu to ensure it is balanced within domain. Attempt to move
6082 * tasks if there is an imbalance.
6084 static int load_balance(int this_cpu, struct rq *this_rq,
6085 struct sched_domain *sd, enum cpu_idle_type idle,
6086 int *continue_balancing)
6088 int ld_moved, cur_ld_moved, active_balance = 0;
6089 struct sched_domain *sd_parent = sd->parent;
6090 struct sched_group *group;
6092 unsigned long flags;
6093 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
6095 struct lb_env env = {
6097 .dst_cpu = this_cpu,
6099 .dst_grpmask = sched_group_cpus(sd->groups),
6101 .loop_break = sched_nr_migrate_break,
6107 * For NEWLY_IDLE load_balancing, we don't need to consider
6108 * other cpus in our group
6110 if (idle == CPU_NEWLY_IDLE)
6111 env.dst_grpmask = NULL;
6113 cpumask_copy(cpus, cpu_active_mask);
6115 schedstat_inc(sd, lb_count[idle]);
6118 if (!should_we_balance(&env)) {
6119 *continue_balancing = 0;
6123 group = find_busiest_group(&env);
6125 schedstat_inc(sd, lb_nobusyg[idle]);
6129 busiest = find_busiest_queue(&env, group);
6131 schedstat_inc(sd, lb_nobusyq[idle]);
6135 BUG_ON(busiest == env.dst_rq);
6137 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6140 if (busiest->nr_running > 1) {
6142 * Attempt to move tasks. If find_busiest_group has found
6143 * an imbalance but busiest->nr_running <= 1, the group is
6144 * still unbalanced. ld_moved simply stays zero, so it is
6145 * correctly treated as an imbalance.
6147 env.flags |= LBF_ALL_PINNED;
6148 env.src_cpu = busiest->cpu;
6149 env.src_rq = busiest;
6150 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6153 local_irq_save(flags);
6154 double_rq_lock(env.dst_rq, busiest);
6157 * cur_ld_moved - load moved in current iteration
6158 * ld_moved - cumulative load moved across iterations
6160 cur_ld_moved = move_tasks(&env);
6161 ld_moved += cur_ld_moved;
6162 double_rq_unlock(env.dst_rq, busiest);
6163 local_irq_restore(flags);
6166 * some other cpu did the load balance for us.
6168 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6169 resched_cpu(env.dst_cpu);
6171 if (env.flags & LBF_NEED_BREAK) {
6172 env.flags &= ~LBF_NEED_BREAK;
6177 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6178 * us and move them to an alternate dst_cpu in our sched_group
6179 * where they can run. The upper limit on how many times we
6180 * iterate on same src_cpu is dependent on number of cpus in our
6183 * This changes load balance semantics a bit on who can move
6184 * load to a given_cpu. In addition to the given_cpu itself
6185 * (or a ilb_cpu acting on its behalf where given_cpu is
6186 * nohz-idle), we now have balance_cpu in a position to move
6187 * load to given_cpu. In rare situations, this may cause
6188 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6189 * _independently_ and at _same_ time to move some load to
6190 * given_cpu) causing exceess load to be moved to given_cpu.
6191 * This however should not happen so much in practice and
6192 * moreover subsequent load balance cycles should correct the
6193 * excess load moved.
6195 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6197 /* Prevent to re-select dst_cpu via env's cpus */
6198 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6200 env.dst_rq = cpu_rq(env.new_dst_cpu);
6201 env.dst_cpu = env.new_dst_cpu;
6202 env.flags &= ~LBF_DST_PINNED;
6204 env.loop_break = sched_nr_migrate_break;
6207 * Go back to "more_balance" rather than "redo" since we
6208 * need to continue with same src_cpu.
6214 * We failed to reach balance because of affinity.
6217 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6219 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6220 *group_imbalance = 1;
6221 } else if (*group_imbalance)
6222 *group_imbalance = 0;
6225 /* All tasks on this runqueue were pinned by CPU affinity */
6226 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6227 cpumask_clear_cpu(cpu_of(busiest), cpus);
6228 if (!cpumask_empty(cpus)) {
6230 env.loop_break = sched_nr_migrate_break;
6238 schedstat_inc(sd, lb_failed[idle]);
6240 * Increment the failure counter only on periodic balance.
6241 * We do not want newidle balance, which can be very
6242 * frequent, pollute the failure counter causing
6243 * excessive cache_hot migrations and active balances.
6245 if (idle != CPU_NEWLY_IDLE)
6246 sd->nr_balance_failed++;
6248 if (need_active_balance(&env)) {
6249 raw_spin_lock_irqsave(&busiest->lock, flags);
6251 /* don't kick the active_load_balance_cpu_stop,
6252 * if the curr task on busiest cpu can't be
6255 if (!cpumask_test_cpu(this_cpu,
6256 tsk_cpus_allowed(busiest->curr))) {
6257 raw_spin_unlock_irqrestore(&busiest->lock,
6259 env.flags |= LBF_ALL_PINNED;
6260 goto out_one_pinned;
6264 * ->active_balance synchronizes accesses to
6265 * ->active_balance_work. Once set, it's cleared
6266 * only after active load balance is finished.
6268 if (!busiest->active_balance) {
6269 busiest->active_balance = 1;
6270 busiest->push_cpu = this_cpu;
6273 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6275 if (active_balance) {
6276 stop_one_cpu_nowait(cpu_of(busiest),
6277 active_load_balance_cpu_stop, busiest,
6278 &busiest->active_balance_work);
6282 * We've kicked active balancing, reset the failure
6285 sd->nr_balance_failed = sd->cache_nice_tries+1;
6288 sd->nr_balance_failed = 0;
6290 if (likely(!active_balance)) {
6291 /* We were unbalanced, so reset the balancing interval */
6292 sd->balance_interval = sd->min_interval;
6295 * If we've begun active balancing, start to back off. This
6296 * case may not be covered by the all_pinned logic if there
6297 * is only 1 task on the busy runqueue (because we don't call
6300 if (sd->balance_interval < sd->max_interval)
6301 sd->balance_interval *= 2;
6307 schedstat_inc(sd, lb_balanced[idle]);
6309 sd->nr_balance_failed = 0;
6312 /* tune up the balancing interval */
6313 if (((env.flags & LBF_ALL_PINNED) &&
6314 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6315 (sd->balance_interval < sd->max_interval))
6316 sd->balance_interval *= 2;
6324 * idle_balance is called by schedule() if this_cpu is about to become
6325 * idle. Attempts to pull tasks from other CPUs.
6327 void idle_balance(int this_cpu, struct rq *this_rq)
6329 struct sched_domain *sd;
6330 int pulled_task = 0;
6331 unsigned long next_balance = jiffies + HZ;
6334 this_rq->idle_stamp = rq_clock(this_rq);
6336 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6340 * Drop the rq->lock, but keep IRQ/preempt disabled.
6342 raw_spin_unlock(&this_rq->lock);
6344 update_blocked_averages(this_cpu);
6346 for_each_domain(this_cpu, sd) {
6347 unsigned long interval;
6348 int continue_balancing = 1;
6349 u64 t0, domain_cost;
6351 if (!(sd->flags & SD_LOAD_BALANCE))
6354 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6357 if (sd->flags & SD_BALANCE_NEWIDLE) {
6358 t0 = sched_clock_cpu(this_cpu);
6360 /* If we've pulled tasks over stop searching: */
6361 pulled_task = load_balance(this_cpu, this_rq,
6363 &continue_balancing);
6365 domain_cost = sched_clock_cpu(this_cpu) - t0;
6366 if (domain_cost > sd->max_newidle_lb_cost)
6367 sd->max_newidle_lb_cost = domain_cost;
6369 curr_cost += domain_cost;
6372 interval = msecs_to_jiffies(sd->balance_interval);
6373 if (time_after(next_balance, sd->last_balance + interval))
6374 next_balance = sd->last_balance + interval;
6376 this_rq->idle_stamp = 0;
6382 raw_spin_lock(&this_rq->lock);
6384 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6386 * We are going idle. next_balance may be set based on
6387 * a busy processor. So reset next_balance.
6389 this_rq->next_balance = next_balance;
6392 if (curr_cost > this_rq->max_idle_balance_cost)
6393 this_rq->max_idle_balance_cost = curr_cost;
6397 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6398 * running tasks off the busiest CPU onto idle CPUs. It requires at
6399 * least 1 task to be running on each physical CPU where possible, and
6400 * avoids physical / logical imbalances.
6402 static int active_load_balance_cpu_stop(void *data)
6404 struct rq *busiest_rq = data;
6405 int busiest_cpu = cpu_of(busiest_rq);
6406 int target_cpu = busiest_rq->push_cpu;
6407 struct rq *target_rq = cpu_rq(target_cpu);
6408 struct sched_domain *sd;
6410 raw_spin_lock_irq(&busiest_rq->lock);
6412 /* make sure the requested cpu hasn't gone down in the meantime */
6413 if (unlikely(busiest_cpu != smp_processor_id() ||
6414 !busiest_rq->active_balance))
6417 /* Is there any task to move? */
6418 if (busiest_rq->nr_running <= 1)
6422 * This condition is "impossible", if it occurs
6423 * we need to fix it. Originally reported by
6424 * Bjorn Helgaas on a 128-cpu setup.
6426 BUG_ON(busiest_rq == target_rq);
6428 /* move a task from busiest_rq to target_rq */
6429 double_lock_balance(busiest_rq, target_rq);
6431 /* Search for an sd spanning us and the target CPU. */
6433 for_each_domain(target_cpu, sd) {
6434 if ((sd->flags & SD_LOAD_BALANCE) &&
6435 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6440 struct lb_env env = {
6442 .dst_cpu = target_cpu,
6443 .dst_rq = target_rq,
6444 .src_cpu = busiest_rq->cpu,
6445 .src_rq = busiest_rq,
6449 schedstat_inc(sd, alb_count);
6451 if (move_one_task(&env))
6452 schedstat_inc(sd, alb_pushed);
6454 schedstat_inc(sd, alb_failed);
6457 double_unlock_balance(busiest_rq, target_rq);
6459 busiest_rq->active_balance = 0;
6460 raw_spin_unlock_irq(&busiest_rq->lock);
6464 #ifdef CONFIG_NO_HZ_COMMON
6466 * idle load balancing details
6467 * - When one of the busy CPUs notice that there may be an idle rebalancing
6468 * needed, they will kick the idle load balancer, which then does idle
6469 * load balancing for all the idle CPUs.
6472 cpumask_var_t idle_cpus_mask;
6474 unsigned long next_balance; /* in jiffy units */
6475 } nohz ____cacheline_aligned;
6477 static inline int find_new_ilb(int call_cpu)
6479 int ilb = cpumask_first(nohz.idle_cpus_mask);
6481 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6488 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6489 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6490 * CPU (if there is one).
6492 static void nohz_balancer_kick(int cpu)
6496 nohz.next_balance++;
6498 ilb_cpu = find_new_ilb(cpu);
6500 if (ilb_cpu >= nr_cpu_ids)
6503 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6506 * Use smp_send_reschedule() instead of resched_cpu().
6507 * This way we generate a sched IPI on the target cpu which
6508 * is idle. And the softirq performing nohz idle load balance
6509 * will be run before returning from the IPI.
6511 smp_send_reschedule(ilb_cpu);
6515 static inline void nohz_balance_exit_idle(int cpu)
6517 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6518 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6519 atomic_dec(&nohz.nr_cpus);
6520 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6524 static inline void set_cpu_sd_state_busy(void)
6526 struct sched_domain *sd;
6529 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6531 if (!sd || !sd->nohz_idle)
6535 for (; sd; sd = sd->parent)
6536 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6541 void set_cpu_sd_state_idle(void)
6543 struct sched_domain *sd;
6546 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6548 if (!sd || sd->nohz_idle)
6552 for (; sd; sd = sd->parent)
6553 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6559 * This routine will record that the cpu is going idle with tick stopped.
6560 * This info will be used in performing idle load balancing in the future.
6562 void nohz_balance_enter_idle(int cpu)
6565 * If this cpu is going down, then nothing needs to be done.
6567 if (!cpu_active(cpu))
6570 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6573 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6574 atomic_inc(&nohz.nr_cpus);
6575 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6578 static int sched_ilb_notifier(struct notifier_block *nfb,
6579 unsigned long action, void *hcpu)
6581 switch (action & ~CPU_TASKS_FROZEN) {
6583 nohz_balance_exit_idle(smp_processor_id());
6591 static DEFINE_SPINLOCK(balancing);
6594 * Scale the max load_balance interval with the number of CPUs in the system.
6595 * This trades load-balance latency on larger machines for less cross talk.
6597 void update_max_interval(void)
6599 max_load_balance_interval = HZ*num_online_cpus()/10;
6603 * It checks each scheduling domain to see if it is due to be balanced,
6604 * and initiates a balancing operation if so.
6606 * Balancing parameters are set up in init_sched_domains.
6608 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6610 int continue_balancing = 1;
6611 struct rq *rq = cpu_rq(cpu);
6612 unsigned long interval;
6613 struct sched_domain *sd;
6614 /* Earliest time when we have to do rebalance again */
6615 unsigned long next_balance = jiffies + 60*HZ;
6616 int update_next_balance = 0;
6617 int need_serialize, need_decay = 0;
6620 update_blocked_averages(cpu);
6623 for_each_domain(cpu, sd) {
6625 * Decay the newidle max times here because this is a regular
6626 * visit to all the domains. Decay ~1% per second.
6628 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6629 sd->max_newidle_lb_cost =
6630 (sd->max_newidle_lb_cost * 253) / 256;
6631 sd->next_decay_max_lb_cost = jiffies + HZ;
6634 max_cost += sd->max_newidle_lb_cost;
6636 if (!(sd->flags & SD_LOAD_BALANCE))
6640 * Stop the load balance at this level. There is another
6641 * CPU in our sched group which is doing load balancing more
6644 if (!continue_balancing) {
6650 interval = sd->balance_interval;
6651 if (idle != CPU_IDLE)
6652 interval *= sd->busy_factor;
6654 /* scale ms to jiffies */
6655 interval = msecs_to_jiffies(interval);
6656 interval = clamp(interval, 1UL, max_load_balance_interval);
6658 need_serialize = sd->flags & SD_SERIALIZE;
6660 if (need_serialize) {
6661 if (!spin_trylock(&balancing))
6665 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6666 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6668 * The LBF_DST_PINNED logic could have changed
6669 * env->dst_cpu, so we can't know our idle
6670 * state even if we migrated tasks. Update it.
6672 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6674 sd->last_balance = jiffies;
6677 spin_unlock(&balancing);
6679 if (time_after(next_balance, sd->last_balance + interval)) {
6680 next_balance = sd->last_balance + interval;
6681 update_next_balance = 1;
6686 * Ensure the rq-wide value also decays but keep it at a
6687 * reasonable floor to avoid funnies with rq->avg_idle.
6689 rq->max_idle_balance_cost =
6690 max((u64)sysctl_sched_migration_cost, max_cost);
6695 * next_balance will be updated only when there is a need.
6696 * When the cpu is attached to null domain for ex, it will not be
6699 if (likely(update_next_balance))
6700 rq->next_balance = next_balance;
6703 #ifdef CONFIG_NO_HZ_COMMON
6705 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6706 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6708 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6710 struct rq *this_rq = cpu_rq(this_cpu);
6714 if (idle != CPU_IDLE ||
6715 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6718 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6719 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6723 * If this cpu gets work to do, stop the load balancing
6724 * work being done for other cpus. Next load
6725 * balancing owner will pick it up.
6730 rq = cpu_rq(balance_cpu);
6732 raw_spin_lock_irq(&rq->lock);
6733 update_rq_clock(rq);
6734 update_idle_cpu_load(rq);
6735 raw_spin_unlock_irq(&rq->lock);
6737 rebalance_domains(balance_cpu, CPU_IDLE);
6739 if (time_after(this_rq->next_balance, rq->next_balance))
6740 this_rq->next_balance = rq->next_balance;
6742 nohz.next_balance = this_rq->next_balance;
6744 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6748 * Current heuristic for kicking the idle load balancer in the presence
6749 * of an idle cpu is the system.
6750 * - This rq has more than one task.
6751 * - At any scheduler domain level, this cpu's scheduler group has multiple
6752 * busy cpu's exceeding the group's power.
6753 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6754 * domain span are idle.
6756 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6758 unsigned long now = jiffies;
6759 struct sched_domain *sd;
6761 if (unlikely(idle_cpu(cpu)))
6765 * We may be recently in ticked or tickless idle mode. At the first
6766 * busy tick after returning from idle, we will update the busy stats.
6768 set_cpu_sd_state_busy();
6769 nohz_balance_exit_idle(cpu);
6772 * None are in tickless mode and hence no need for NOHZ idle load
6775 if (likely(!atomic_read(&nohz.nr_cpus)))
6778 if (time_before(now, nohz.next_balance))
6781 if (rq->nr_running >= 2)
6785 for_each_domain(cpu, sd) {
6786 struct sched_group *sg = sd->groups;
6787 struct sched_group_power *sgp = sg->sgp;
6788 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6790 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6791 goto need_kick_unlock;
6793 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6794 && (cpumask_first_and(nohz.idle_cpus_mask,
6795 sched_domain_span(sd)) < cpu))
6796 goto need_kick_unlock;
6798 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6810 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6814 * run_rebalance_domains is triggered when needed from the scheduler tick.
6815 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6817 static void run_rebalance_domains(struct softirq_action *h)
6819 int this_cpu = smp_processor_id();
6820 struct rq *this_rq = cpu_rq(this_cpu);
6821 enum cpu_idle_type idle = this_rq->idle_balance ?
6822 CPU_IDLE : CPU_NOT_IDLE;
6824 rebalance_domains(this_cpu, idle);
6827 * If this cpu has a pending nohz_balance_kick, then do the
6828 * balancing on behalf of the other idle cpus whose ticks are
6831 nohz_idle_balance(this_cpu, idle);
6834 static inline int on_null_domain(int cpu)
6836 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6840 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6842 void trigger_load_balance(struct rq *rq, int cpu)
6844 /* Don't need to rebalance while attached to NULL domain */
6845 if (time_after_eq(jiffies, rq->next_balance) &&
6846 likely(!on_null_domain(cpu)))
6847 raise_softirq(SCHED_SOFTIRQ);
6848 #ifdef CONFIG_NO_HZ_COMMON
6849 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6850 nohz_balancer_kick(cpu);
6854 static void rq_online_fair(struct rq *rq)
6859 static void rq_offline_fair(struct rq *rq)
6863 /* Ensure any throttled groups are reachable by pick_next_task */
6864 unthrottle_offline_cfs_rqs(rq);
6867 #endif /* CONFIG_SMP */
6870 * scheduler tick hitting a task of our scheduling class:
6872 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6874 struct cfs_rq *cfs_rq;
6875 struct sched_entity *se = &curr->se;
6877 for_each_sched_entity(se) {
6878 cfs_rq = cfs_rq_of(se);
6879 entity_tick(cfs_rq, se, queued);
6882 if (numabalancing_enabled)
6883 task_tick_numa(rq, curr);
6885 update_rq_runnable_avg(rq, 1);
6889 * called on fork with the child task as argument from the parent's context
6890 * - child not yet on the tasklist
6891 * - preemption disabled
6893 static void task_fork_fair(struct task_struct *p)
6895 struct cfs_rq *cfs_rq;
6896 struct sched_entity *se = &p->se, *curr;
6897 int this_cpu = smp_processor_id();
6898 struct rq *rq = this_rq();
6899 unsigned long flags;
6901 raw_spin_lock_irqsave(&rq->lock, flags);
6903 update_rq_clock(rq);
6905 cfs_rq = task_cfs_rq(current);
6906 curr = cfs_rq->curr;
6909 * Not only the cpu but also the task_group of the parent might have
6910 * been changed after parent->se.parent,cfs_rq were copied to
6911 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6912 * of child point to valid ones.
6915 __set_task_cpu(p, this_cpu);
6918 update_curr(cfs_rq);
6921 se->vruntime = curr->vruntime;
6922 place_entity(cfs_rq, se, 1);
6924 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6926 * Upon rescheduling, sched_class::put_prev_task() will place
6927 * 'current' within the tree based on its new key value.
6929 swap(curr->vruntime, se->vruntime);
6930 resched_task(rq->curr);
6933 se->vruntime -= cfs_rq->min_vruntime;
6935 raw_spin_unlock_irqrestore(&rq->lock, flags);
6939 * Priority of the task has changed. Check to see if we preempt
6943 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6949 * Reschedule if we are currently running on this runqueue and
6950 * our priority decreased, or if we are not currently running on
6951 * this runqueue and our priority is higher than the current's
6953 if (rq->curr == p) {
6954 if (p->prio > oldprio)
6955 resched_task(rq->curr);
6957 check_preempt_curr(rq, p, 0);
6960 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6962 struct sched_entity *se = &p->se;
6963 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6966 * Ensure the task's vruntime is normalized, so that when its
6967 * switched back to the fair class the enqueue_entity(.flags=0) will
6968 * do the right thing.
6970 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6971 * have normalized the vruntime, if it was !on_rq, then only when
6972 * the task is sleeping will it still have non-normalized vruntime.
6974 if (!se->on_rq && p->state != TASK_RUNNING) {
6976 * Fix up our vruntime so that the current sleep doesn't
6977 * cause 'unlimited' sleep bonus.
6979 place_entity(cfs_rq, se, 0);
6980 se->vruntime -= cfs_rq->min_vruntime;
6985 * Remove our load from contribution when we leave sched_fair
6986 * and ensure we don't carry in an old decay_count if we
6989 if (se->avg.decay_count) {
6990 __synchronize_entity_decay(se);
6991 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6997 * We switched to the sched_fair class.
6999 static void switched_to_fair(struct rq *rq, struct task_struct *p)
7005 * We were most likely switched from sched_rt, so
7006 * kick off the schedule if running, otherwise just see
7007 * if we can still preempt the current task.
7010 resched_task(rq->curr);
7012 check_preempt_curr(rq, p, 0);
7015 /* Account for a task changing its policy or group.
7017 * This routine is mostly called to set cfs_rq->curr field when a task
7018 * migrates between groups/classes.
7020 static void set_curr_task_fair(struct rq *rq)
7022 struct sched_entity *se = &rq->curr->se;
7024 for_each_sched_entity(se) {
7025 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7027 set_next_entity(cfs_rq, se);
7028 /* ensure bandwidth has been allocated on our new cfs_rq */
7029 account_cfs_rq_runtime(cfs_rq, 0);
7033 void init_cfs_rq(struct cfs_rq *cfs_rq)
7035 cfs_rq->tasks_timeline = RB_ROOT;
7036 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7037 #ifndef CONFIG_64BIT
7038 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7041 atomic64_set(&cfs_rq->decay_counter, 1);
7042 atomic_long_set(&cfs_rq->removed_load, 0);
7046 #ifdef CONFIG_FAIR_GROUP_SCHED
7047 static void task_move_group_fair(struct task_struct *p, int on_rq)
7049 struct cfs_rq *cfs_rq;
7051 * If the task was not on the rq at the time of this cgroup movement
7052 * it must have been asleep, sleeping tasks keep their ->vruntime
7053 * absolute on their old rq until wakeup (needed for the fair sleeper
7054 * bonus in place_entity()).
7056 * If it was on the rq, we've just 'preempted' it, which does convert
7057 * ->vruntime to a relative base.
7059 * Make sure both cases convert their relative position when migrating
7060 * to another cgroup's rq. This does somewhat interfere with the
7061 * fair sleeper stuff for the first placement, but who cares.
7064 * When !on_rq, vruntime of the task has usually NOT been normalized.
7065 * But there are some cases where it has already been normalized:
7067 * - Moving a forked child which is waiting for being woken up by
7068 * wake_up_new_task().
7069 * - Moving a task which has been woken up by try_to_wake_up() and
7070 * waiting for actually being woken up by sched_ttwu_pending().
7072 * To prevent boost or penalty in the new cfs_rq caused by delta
7073 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
7075 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
7079 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
7080 set_task_rq(p, task_cpu(p));
7082 cfs_rq = cfs_rq_of(&p->se);
7083 p->se.vruntime += cfs_rq->min_vruntime;
7086 * migrate_task_rq_fair() will have removed our previous
7087 * contribution, but we must synchronize for ongoing future
7090 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
7091 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
7096 void free_fair_sched_group(struct task_group *tg)
7100 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
7102 for_each_possible_cpu(i) {
7104 kfree(tg->cfs_rq[i]);
7113 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7115 struct cfs_rq *cfs_rq;
7116 struct sched_entity *se;
7119 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7122 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7126 tg->shares = NICE_0_LOAD;
7128 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
7130 for_each_possible_cpu(i) {
7131 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7132 GFP_KERNEL, cpu_to_node(i));
7136 se = kzalloc_node(sizeof(struct sched_entity),
7137 GFP_KERNEL, cpu_to_node(i));
7141 init_cfs_rq(cfs_rq);
7142 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7153 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7155 struct rq *rq = cpu_rq(cpu);
7156 unsigned long flags;
7159 * Only empty task groups can be destroyed; so we can speculatively
7160 * check on_list without danger of it being re-added.
7162 if (!tg->cfs_rq[cpu]->on_list)
7165 raw_spin_lock_irqsave(&rq->lock, flags);
7166 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7167 raw_spin_unlock_irqrestore(&rq->lock, flags);
7170 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7171 struct sched_entity *se, int cpu,
7172 struct sched_entity *parent)
7174 struct rq *rq = cpu_rq(cpu);
7178 init_cfs_rq_runtime(cfs_rq);
7180 tg->cfs_rq[cpu] = cfs_rq;
7183 /* se could be NULL for root_task_group */
7188 se->cfs_rq = &rq->cfs;
7190 se->cfs_rq = parent->my_q;
7193 update_load_set(&se->load, 0);
7194 se->parent = parent;
7197 static DEFINE_MUTEX(shares_mutex);
7199 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7202 unsigned long flags;
7205 * We can't change the weight of the root cgroup.
7210 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7212 mutex_lock(&shares_mutex);
7213 if (tg->shares == shares)
7216 tg->shares = shares;
7217 for_each_possible_cpu(i) {
7218 struct rq *rq = cpu_rq(i);
7219 struct sched_entity *se;
7222 /* Propagate contribution to hierarchy */
7223 raw_spin_lock_irqsave(&rq->lock, flags);
7225 /* Possible calls to update_curr() need rq clock */
7226 update_rq_clock(rq);
7227 for_each_sched_entity(se)
7228 update_cfs_shares(group_cfs_rq(se));
7229 raw_spin_unlock_irqrestore(&rq->lock, flags);
7233 mutex_unlock(&shares_mutex);
7236 #else /* CONFIG_FAIR_GROUP_SCHED */
7238 void free_fair_sched_group(struct task_group *tg) { }
7240 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7245 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7247 #endif /* CONFIG_FAIR_GROUP_SCHED */
7250 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7252 struct sched_entity *se = &task->se;
7253 unsigned int rr_interval = 0;
7256 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7259 if (rq->cfs.load.weight)
7260 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7266 * All the scheduling class methods:
7268 const struct sched_class fair_sched_class = {
7269 .next = &idle_sched_class,
7270 .enqueue_task = enqueue_task_fair,
7271 .dequeue_task = dequeue_task_fair,
7272 .yield_task = yield_task_fair,
7273 .yield_to_task = yield_to_task_fair,
7275 .check_preempt_curr = check_preempt_wakeup,
7277 .pick_next_task = pick_next_task_fair,
7278 .put_prev_task = put_prev_task_fair,
7281 .select_task_rq = select_task_rq_fair,
7282 .migrate_task_rq = migrate_task_rq_fair,
7284 .rq_online = rq_online_fair,
7285 .rq_offline = rq_offline_fair,
7287 .task_waking = task_waking_fair,
7290 .set_curr_task = set_curr_task_fair,
7291 .task_tick = task_tick_fair,
7292 .task_fork = task_fork_fair,
7294 .prio_changed = prio_changed_fair,
7295 .switched_from = switched_from_fair,
7296 .switched_to = switched_to_fair,
7298 .get_rr_interval = get_rr_interval_fair,
7300 #ifdef CONFIG_FAIR_GROUP_SCHED
7301 .task_move_group = task_move_group_fair,
7305 #ifdef CONFIG_SCHED_DEBUG
7306 void print_cfs_stats(struct seq_file *m, int cpu)
7308 struct cfs_rq *cfs_rq;
7311 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7312 print_cfs_rq(m, cpu, cfs_rq);
7317 __init void init_sched_fair_class(void)
7320 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7322 #ifdef CONFIG_NO_HZ_COMMON
7323 nohz.next_balance = jiffies;
7324 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7325 cpu_notifier(sched_ilb_notifier, 0);