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;
829 unsigned int sysctl_numa_balancing_scan_period_reset = 60000;
831 /* Portion of address space to scan in MB */
832 unsigned int sysctl_numa_balancing_scan_size = 256;
834 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
835 unsigned int sysctl_numa_balancing_scan_delay = 1000;
837 static unsigned int task_nr_scan_windows(struct task_struct *p)
839 unsigned long rss = 0;
840 unsigned long nr_scan_pages;
843 * Calculations based on RSS as non-present and empty pages are skipped
844 * by the PTE scanner and NUMA hinting faults should be trapped based
847 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
848 rss = get_mm_rss(p->mm);
852 rss = round_up(rss, nr_scan_pages);
853 return rss / nr_scan_pages;
856 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
857 #define MAX_SCAN_WINDOW 2560
859 static unsigned int task_scan_min(struct task_struct *p)
861 unsigned int scan, floor;
862 unsigned int windows = 1;
864 if (sysctl_numa_balancing_scan_size < MAX_SCAN_WINDOW)
865 windows = MAX_SCAN_WINDOW / sysctl_numa_balancing_scan_size;
866 floor = 1000 / windows;
868 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
869 return max_t(unsigned int, floor, scan);
872 static unsigned int task_scan_max(struct task_struct *p)
874 unsigned int smin = task_scan_min(p);
877 /* Watch for min being lower than max due to floor calculations */
878 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
879 return max(smin, smax);
883 * Once a preferred node is selected the scheduler balancer will prefer moving
884 * a task to that node for sysctl_numa_balancing_settle_count number of PTE
885 * scans. This will give the process the chance to accumulate more faults on
886 * the preferred node but still allow the scheduler to move the task again if
887 * the nodes CPUs are overloaded.
889 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
894 spinlock_t lock; /* nr_tasks, tasks */
897 struct list_head task_list;
900 atomic_long_t total_faults;
901 atomic_long_t faults[0];
904 pid_t task_numa_group_id(struct task_struct *p)
906 return p->numa_group ? p->numa_group->gid : 0;
909 static inline int task_faults_idx(int nid, int priv)
911 return 2 * nid + priv;
914 static inline unsigned long task_faults(struct task_struct *p, int nid)
919 return p->numa_faults[task_faults_idx(nid, 0)] +
920 p->numa_faults[task_faults_idx(nid, 1)];
923 static inline unsigned long group_faults(struct task_struct *p, int nid)
928 return atomic_long_read(&p->numa_group->faults[2*nid]) +
929 atomic_long_read(&p->numa_group->faults[2*nid+1]);
933 * These return the fraction of accesses done by a particular task, or
934 * task group, on a particular numa node. The group weight is given a
935 * larger multiplier, in order to group tasks together that are almost
936 * evenly spread out between numa nodes.
938 static inline unsigned long task_weight(struct task_struct *p, int nid)
940 unsigned long total_faults;
945 total_faults = p->total_numa_faults;
950 return 1000 * task_faults(p, nid) / total_faults;
953 static inline unsigned long group_weight(struct task_struct *p, int nid)
955 unsigned long total_faults;
960 total_faults = atomic_long_read(&p->numa_group->total_faults);
965 return 1000 * group_faults(p, nid) / total_faults;
968 static unsigned long weighted_cpuload(const int cpu);
969 static unsigned long source_load(int cpu, int type);
970 static unsigned long target_load(int cpu, int type);
971 static unsigned long power_of(int cpu);
972 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
974 /* Cached statistics for all CPUs within a node */
976 unsigned long nr_running;
979 /* Total compute capacity of CPUs on a node */
982 /* Approximate capacity in terms of runnable tasks on a node */
983 unsigned long capacity;
988 * XXX borrowed from update_sg_lb_stats
990 static void update_numa_stats(struct numa_stats *ns, int nid)
994 memset(ns, 0, sizeof(*ns));
995 for_each_cpu(cpu, cpumask_of_node(nid)) {
996 struct rq *rq = cpu_rq(cpu);
998 ns->nr_running += rq->nr_running;
999 ns->load += weighted_cpuload(cpu);
1000 ns->power += power_of(cpu);
1003 ns->load = (ns->load * SCHED_POWER_SCALE) / ns->power;
1004 ns->capacity = DIV_ROUND_CLOSEST(ns->power, SCHED_POWER_SCALE);
1005 ns->has_capacity = (ns->nr_running < ns->capacity);
1008 struct task_numa_env {
1009 struct task_struct *p;
1011 int src_cpu, src_nid;
1012 int dst_cpu, dst_nid;
1014 struct numa_stats src_stats, dst_stats;
1016 int imbalance_pct, idx;
1018 struct task_struct *best_task;
1023 static void task_numa_assign(struct task_numa_env *env,
1024 struct task_struct *p, long imp)
1027 put_task_struct(env->best_task);
1032 env->best_imp = imp;
1033 env->best_cpu = env->dst_cpu;
1037 * This checks if the overall compute and NUMA accesses of the system would
1038 * be improved if the source tasks was migrated to the target dst_cpu taking
1039 * into account that it might be best if task running on the dst_cpu should
1040 * be exchanged with the source task
1042 static void task_numa_compare(struct task_numa_env *env,
1043 long taskimp, long groupimp)
1045 struct rq *src_rq = cpu_rq(env->src_cpu);
1046 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1047 struct task_struct *cur;
1048 long dst_load, src_load;
1050 long imp = (groupimp > 0) ? groupimp : taskimp;
1053 cur = ACCESS_ONCE(dst_rq->curr);
1054 if (cur->pid == 0) /* idle */
1058 * "imp" is the fault differential for the source task between the
1059 * source and destination node. Calculate the total differential for
1060 * the source task and potential destination task. The more negative
1061 * the value is, the more rmeote accesses that would be expected to
1062 * be incurred if the tasks were swapped.
1065 /* Skip this swap candidate if cannot move to the source cpu */
1066 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1070 * If dst and source tasks are in the same NUMA group, or not
1071 * in any group then look only at task weights.
1073 if (cur->numa_group == env->p->numa_group) {
1074 imp = taskimp + task_weight(cur, env->src_nid) -
1075 task_weight(cur, env->dst_nid);
1077 * Add some hysteresis to prevent swapping the
1078 * tasks within a group over tiny differences.
1080 if (cur->numa_group)
1084 * Compare the group weights. If a task is all by
1085 * itself (not part of a group), use the task weight
1088 if (env->p->numa_group)
1093 if (cur->numa_group)
1094 imp += group_weight(cur, env->src_nid) -
1095 group_weight(cur, env->dst_nid);
1097 imp += task_weight(cur, env->src_nid) -
1098 task_weight(cur, env->dst_nid);
1102 if (imp < env->best_imp)
1106 /* Is there capacity at our destination? */
1107 if (env->src_stats.has_capacity &&
1108 !env->dst_stats.has_capacity)
1114 /* Balance doesn't matter much if we're running a task per cpu */
1115 if (src_rq->nr_running == 1 && dst_rq->nr_running == 1)
1119 * In the overloaded case, try and keep the load balanced.
1122 dst_load = env->dst_stats.load;
1123 src_load = env->src_stats.load;
1125 /* XXX missing power terms */
1126 load = task_h_load(env->p);
1131 load = task_h_load(cur);
1136 /* make src_load the smaller */
1137 if (dst_load < src_load)
1138 swap(dst_load, src_load);
1140 if (src_load * env->imbalance_pct < dst_load * 100)
1144 task_numa_assign(env, cur, imp);
1149 static void task_numa_find_cpu(struct task_numa_env *env,
1150 long taskimp, long groupimp)
1154 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1155 /* Skip this CPU if the source task cannot migrate */
1156 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1160 task_numa_compare(env, taskimp, groupimp);
1164 static int task_numa_migrate(struct task_struct *p)
1166 struct task_numa_env env = {
1169 .src_cpu = task_cpu(p),
1170 .src_nid = task_node(p),
1172 .imbalance_pct = 112,
1178 struct sched_domain *sd;
1179 unsigned long taskweight, groupweight;
1181 long taskimp, groupimp;
1184 * Pick the lowest SD_NUMA domain, as that would have the smallest
1185 * imbalance and would be the first to start moving tasks about.
1187 * And we want to avoid any moving of tasks about, as that would create
1188 * random movement of tasks -- counter the numa conditions we're trying
1192 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1193 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1196 taskweight = task_weight(p, env.src_nid);
1197 groupweight = group_weight(p, env.src_nid);
1198 update_numa_stats(&env.src_stats, env.src_nid);
1199 env.dst_nid = p->numa_preferred_nid;
1200 taskimp = task_weight(p, env.dst_nid) - taskweight;
1201 groupimp = group_weight(p, env.dst_nid) - groupweight;
1202 update_numa_stats(&env.dst_stats, env.dst_nid);
1204 /* If the preferred nid has capacity, try to use it. */
1205 if (env.dst_stats.has_capacity)
1206 task_numa_find_cpu(&env, taskimp, groupimp);
1208 /* No space available on the preferred nid. Look elsewhere. */
1209 if (env.best_cpu == -1) {
1210 for_each_online_node(nid) {
1211 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1214 /* Only consider nodes where both task and groups benefit */
1215 taskimp = task_weight(p, nid) - taskweight;
1216 groupimp = group_weight(p, nid) - groupweight;
1217 if (taskimp < 0 && groupimp < 0)
1221 update_numa_stats(&env.dst_stats, env.dst_nid);
1222 task_numa_find_cpu(&env, taskimp, groupimp);
1226 /* No better CPU than the current one was found. */
1227 if (env.best_cpu == -1)
1230 if (env.best_task == NULL) {
1231 int ret = migrate_task_to(p, env.best_cpu);
1235 ret = migrate_swap(p, env.best_task);
1236 put_task_struct(env.best_task);
1240 /* Attempt to migrate a task to a CPU on the preferred node. */
1241 static void numa_migrate_preferred(struct task_struct *p)
1243 /* Success if task is already running on preferred CPU */
1244 p->numa_migrate_retry = 0;
1245 if (cpu_to_node(task_cpu(p)) == p->numa_preferred_nid) {
1247 * If migration is temporarily disabled due to a task migration
1248 * then re-enable it now as the task is running on its
1249 * preferred node and memory should migrate locally
1251 if (!p->numa_migrate_seq)
1252 p->numa_migrate_seq++;
1256 /* This task has no NUMA fault statistics yet */
1257 if (unlikely(p->numa_preferred_nid == -1))
1260 /* Otherwise, try migrate to a CPU on the preferred node */
1261 if (task_numa_migrate(p) != 0)
1262 p->numa_migrate_retry = jiffies + HZ*5;
1265 static void task_numa_placement(struct task_struct *p)
1267 int seq, nid, max_nid = -1, max_group_nid = -1;
1268 unsigned long max_faults = 0, max_group_faults = 0;
1269 spinlock_t *group_lock = NULL;
1271 seq = ACCESS_ONCE(p->mm->numa_scan_seq);
1272 if (p->numa_scan_seq == seq)
1274 p->numa_scan_seq = seq;
1275 p->numa_migrate_seq++;
1276 p->numa_scan_period_max = task_scan_max(p);
1278 /* If the task is part of a group prevent parallel updates to group stats */
1279 if (p->numa_group) {
1280 group_lock = &p->numa_group->lock;
1281 spin_lock(group_lock);
1284 /* Find the node with the highest number of faults */
1285 for_each_online_node(nid) {
1286 unsigned long faults = 0, group_faults = 0;
1289 for (priv = 0; priv < 2; priv++) {
1292 i = task_faults_idx(nid, priv);
1293 diff = -p->numa_faults[i];
1295 /* Decay existing window, copy faults since last scan */
1296 p->numa_faults[i] >>= 1;
1297 p->numa_faults[i] += p->numa_faults_buffer[i];
1298 p->numa_faults_buffer[i] = 0;
1300 faults += p->numa_faults[i];
1301 diff += p->numa_faults[i];
1302 p->total_numa_faults += diff;
1303 if (p->numa_group) {
1304 /* safe because we can only change our own group */
1305 atomic_long_add(diff, &p->numa_group->faults[i]);
1306 atomic_long_add(diff, &p->numa_group->total_faults);
1307 group_faults += atomic_long_read(&p->numa_group->faults[i]);
1311 if (faults > max_faults) {
1312 max_faults = faults;
1316 if (group_faults > max_group_faults) {
1317 max_group_faults = group_faults;
1318 max_group_nid = nid;
1322 if (p->numa_group) {
1324 * If the preferred task and group nids are different,
1325 * iterate over the nodes again to find the best place.
1327 if (max_nid != max_group_nid) {
1328 unsigned long weight, max_weight = 0;
1330 for_each_online_node(nid) {
1331 weight = task_weight(p, nid) + group_weight(p, nid);
1332 if (weight > max_weight) {
1333 max_weight = weight;
1339 spin_unlock(group_lock);
1342 /* Preferred node as the node with the most faults */
1343 if (max_faults && max_nid != p->numa_preferred_nid) {
1344 /* Update the preferred nid and migrate task if possible */
1345 p->numa_preferred_nid = max_nid;
1346 p->numa_migrate_seq = 1;
1347 numa_migrate_preferred(p);
1351 static inline int get_numa_group(struct numa_group *grp)
1353 return atomic_inc_not_zero(&grp->refcount);
1356 static inline void put_numa_group(struct numa_group *grp)
1358 if (atomic_dec_and_test(&grp->refcount))
1359 kfree_rcu(grp, rcu);
1362 static void double_lock(spinlock_t *l1, spinlock_t *l2)
1368 spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
1371 static void task_numa_group(struct task_struct *p, int cpupid)
1373 struct numa_group *grp, *my_grp;
1374 struct task_struct *tsk;
1376 int cpu = cpupid_to_cpu(cpupid);
1379 if (unlikely(!p->numa_group)) {
1380 unsigned int size = sizeof(struct numa_group) +
1381 2*nr_node_ids*sizeof(atomic_long_t);
1383 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1387 atomic_set(&grp->refcount, 1);
1388 spin_lock_init(&grp->lock);
1389 INIT_LIST_HEAD(&grp->task_list);
1392 for (i = 0; i < 2*nr_node_ids; i++)
1393 atomic_long_set(&grp->faults[i], p->numa_faults[i]);
1395 atomic_long_set(&grp->total_faults, p->total_numa_faults);
1397 list_add(&p->numa_entry, &grp->task_list);
1399 rcu_assign_pointer(p->numa_group, grp);
1403 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
1405 if (!cpupid_match_pid(tsk, cpupid))
1408 grp = rcu_dereference(tsk->numa_group);
1412 my_grp = p->numa_group;
1417 * Only join the other group if its bigger; if we're the bigger group,
1418 * the other task will join us.
1420 if (my_grp->nr_tasks > grp->nr_tasks)
1424 * Tie-break on the grp address.
1426 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
1429 if (!get_numa_group(grp))
1440 for (i = 0; i < 2*nr_node_ids; i++) {
1441 atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
1442 atomic_long_add(p->numa_faults[i], &grp->faults[i]);
1444 atomic_long_sub(p->total_numa_faults, &my_grp->total_faults);
1445 atomic_long_add(p->total_numa_faults, &grp->total_faults);
1447 double_lock(&my_grp->lock, &grp->lock);
1449 list_move(&p->numa_entry, &grp->task_list);
1453 spin_unlock(&my_grp->lock);
1454 spin_unlock(&grp->lock);
1456 rcu_assign_pointer(p->numa_group, grp);
1458 put_numa_group(my_grp);
1461 void task_numa_free(struct task_struct *p)
1463 struct numa_group *grp = p->numa_group;
1465 void *numa_faults = p->numa_faults;
1468 for (i = 0; i < 2*nr_node_ids; i++)
1469 atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
1471 atomic_long_sub(p->total_numa_faults, &grp->total_faults);
1473 spin_lock(&grp->lock);
1474 list_del(&p->numa_entry);
1476 spin_unlock(&grp->lock);
1477 rcu_assign_pointer(p->numa_group, NULL);
1478 put_numa_group(grp);
1481 p->numa_faults = NULL;
1482 p->numa_faults_buffer = NULL;
1487 * Got a PROT_NONE fault for a page on @node.
1489 void task_numa_fault(int last_cpupid, int node, int pages, int flags)
1491 struct task_struct *p = current;
1492 bool migrated = flags & TNF_MIGRATED;
1495 if (!numabalancing_enabled)
1498 /* for example, ksmd faulting in a user's mm */
1502 /* Do not worry about placement if exiting */
1503 if (p->state == TASK_DEAD)
1506 /* Allocate buffer to track faults on a per-node basis */
1507 if (unlikely(!p->numa_faults)) {
1508 int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
1510 /* numa_faults and numa_faults_buffer share the allocation */
1511 p->numa_faults = kzalloc(size * 2, GFP_KERNEL|__GFP_NOWARN);
1512 if (!p->numa_faults)
1515 BUG_ON(p->numa_faults_buffer);
1516 p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
1517 p->total_numa_faults = 0;
1521 * First accesses are treated as private, otherwise consider accesses
1522 * to be private if the accessing pid has not changed
1524 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
1527 priv = cpupid_match_pid(p, last_cpupid);
1528 if (!priv && !(flags & TNF_NO_GROUP))
1529 task_numa_group(p, last_cpupid);
1533 * If pages are properly placed (did not migrate) then scan slower.
1534 * This is reset periodically in case of phase changes
1537 /* Initialise if necessary */
1538 if (!p->numa_scan_period_max)
1539 p->numa_scan_period_max = task_scan_max(p);
1541 p->numa_scan_period = min(p->numa_scan_period_max,
1542 p->numa_scan_period + 10);
1545 task_numa_placement(p);
1547 /* Retry task to preferred node migration if it previously failed */
1548 if (p->numa_migrate_retry && time_after(jiffies, p->numa_migrate_retry))
1549 numa_migrate_preferred(p);
1552 p->numa_pages_migrated += pages;
1554 p->numa_faults_buffer[task_faults_idx(node, priv)] += pages;
1557 static void reset_ptenuma_scan(struct task_struct *p)
1559 ACCESS_ONCE(p->mm->numa_scan_seq)++;
1560 p->mm->numa_scan_offset = 0;
1564 * The expensive part of numa migration is done from task_work context.
1565 * Triggered from task_tick_numa().
1567 void task_numa_work(struct callback_head *work)
1569 unsigned long migrate, next_scan, now = jiffies;
1570 struct task_struct *p = current;
1571 struct mm_struct *mm = p->mm;
1572 struct vm_area_struct *vma;
1573 unsigned long start, end;
1574 unsigned long nr_pte_updates = 0;
1577 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
1579 work->next = work; /* protect against double add */
1581 * Who cares about NUMA placement when they're dying.
1583 * NOTE: make sure not to dereference p->mm before this check,
1584 * exit_task_work() happens _after_ exit_mm() so we could be called
1585 * without p->mm even though we still had it when we enqueued this
1588 if (p->flags & PF_EXITING)
1591 if (!mm->numa_next_reset || !mm->numa_next_scan) {
1592 mm->numa_next_scan = now +
1593 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1594 mm->numa_next_reset = now +
1595 msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1599 * Reset the scan period if enough time has gone by. Objective is that
1600 * scanning will be reduced if pages are properly placed. As tasks
1601 * can enter different phases this needs to be re-examined. Lacking
1602 * proper tracking of reference behaviour, this blunt hammer is used.
1604 migrate = mm->numa_next_reset;
1605 if (time_after(now, migrate)) {
1606 p->numa_scan_period = task_scan_min(p);
1607 next_scan = now + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1608 xchg(&mm->numa_next_reset, next_scan);
1612 * Enforce maximal scan/migration frequency..
1614 migrate = mm->numa_next_scan;
1615 if (time_before(now, migrate))
1618 if (p->numa_scan_period == 0) {
1619 p->numa_scan_period_max = task_scan_max(p);
1620 p->numa_scan_period = task_scan_min(p);
1623 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1624 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
1628 * Delay this task enough that another task of this mm will likely win
1629 * the next time around.
1631 p->node_stamp += 2 * TICK_NSEC;
1633 start = mm->numa_scan_offset;
1634 pages = sysctl_numa_balancing_scan_size;
1635 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
1639 down_read(&mm->mmap_sem);
1640 vma = find_vma(mm, start);
1642 reset_ptenuma_scan(p);
1646 for (; vma; vma = vma->vm_next) {
1647 if (!vma_migratable(vma) || !vma_policy_mof(p, vma))
1651 * Shared library pages mapped by multiple processes are not
1652 * migrated as it is expected they are cache replicated. Avoid
1653 * hinting faults in read-only file-backed mappings or the vdso
1654 * as migrating the pages will be of marginal benefit.
1657 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
1661 start = max(start, vma->vm_start);
1662 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
1663 end = min(end, vma->vm_end);
1664 nr_pte_updates += change_prot_numa(vma, start, end);
1667 * Scan sysctl_numa_balancing_scan_size but ensure that
1668 * at least one PTE is updated so that unused virtual
1669 * address space is quickly skipped.
1672 pages -= (end - start) >> PAGE_SHIFT;
1677 } while (end != vma->vm_end);
1682 * If the whole process was scanned without updates then no NUMA
1683 * hinting faults are being recorded and scan rate should be lower.
1685 if (mm->numa_scan_offset == 0 && !nr_pte_updates) {
1686 p->numa_scan_period = min(p->numa_scan_period_max,
1687 p->numa_scan_period << 1);
1689 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
1690 mm->numa_next_scan = next_scan;
1694 * It is possible to reach the end of the VMA list but the last few
1695 * VMAs are not guaranteed to the vma_migratable. If they are not, we
1696 * would find the !migratable VMA on the next scan but not reset the
1697 * scanner to the start so check it now.
1700 mm->numa_scan_offset = start;
1702 reset_ptenuma_scan(p);
1703 up_read(&mm->mmap_sem);
1707 * Drive the periodic memory faults..
1709 void task_tick_numa(struct rq *rq, struct task_struct *curr)
1711 struct callback_head *work = &curr->numa_work;
1715 * We don't care about NUMA placement if we don't have memory.
1717 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
1721 * Using runtime rather than walltime has the dual advantage that
1722 * we (mostly) drive the selection from busy threads and that the
1723 * task needs to have done some actual work before we bother with
1726 now = curr->se.sum_exec_runtime;
1727 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
1729 if (now - curr->node_stamp > period) {
1730 if (!curr->node_stamp)
1731 curr->numa_scan_period = task_scan_min(curr);
1732 curr->node_stamp += period;
1734 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
1735 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
1736 task_work_add(curr, work, true);
1741 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
1744 #endif /* CONFIG_NUMA_BALANCING */
1747 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1749 update_load_add(&cfs_rq->load, se->load.weight);
1750 if (!parent_entity(se))
1751 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
1753 if (entity_is_task(se))
1754 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
1756 cfs_rq->nr_running++;
1760 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
1762 update_load_sub(&cfs_rq->load, se->load.weight);
1763 if (!parent_entity(se))
1764 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
1765 if (entity_is_task(se))
1766 list_del_init(&se->group_node);
1767 cfs_rq->nr_running--;
1770 #ifdef CONFIG_FAIR_GROUP_SCHED
1772 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
1777 * Use this CPU's actual weight instead of the last load_contribution
1778 * to gain a more accurate current total weight. See
1779 * update_cfs_rq_load_contribution().
1781 tg_weight = atomic_long_read(&tg->load_avg);
1782 tg_weight -= cfs_rq->tg_load_contrib;
1783 tg_weight += cfs_rq->load.weight;
1788 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1790 long tg_weight, load, shares;
1792 tg_weight = calc_tg_weight(tg, cfs_rq);
1793 load = cfs_rq->load.weight;
1795 shares = (tg->shares * load);
1797 shares /= tg_weight;
1799 if (shares < MIN_SHARES)
1800 shares = MIN_SHARES;
1801 if (shares > tg->shares)
1802 shares = tg->shares;
1806 # else /* CONFIG_SMP */
1807 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
1811 # endif /* CONFIG_SMP */
1812 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
1813 unsigned long weight)
1816 /* commit outstanding execution time */
1817 if (cfs_rq->curr == se)
1818 update_curr(cfs_rq);
1819 account_entity_dequeue(cfs_rq, se);
1822 update_load_set(&se->load, weight);
1825 account_entity_enqueue(cfs_rq, se);
1828 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
1830 static void update_cfs_shares(struct cfs_rq *cfs_rq)
1832 struct task_group *tg;
1833 struct sched_entity *se;
1837 se = tg->se[cpu_of(rq_of(cfs_rq))];
1838 if (!se || throttled_hierarchy(cfs_rq))
1841 if (likely(se->load.weight == tg->shares))
1844 shares = calc_cfs_shares(cfs_rq, tg);
1846 reweight_entity(cfs_rq_of(se), se, shares);
1848 #else /* CONFIG_FAIR_GROUP_SCHED */
1849 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
1852 #endif /* CONFIG_FAIR_GROUP_SCHED */
1856 * We choose a half-life close to 1 scheduling period.
1857 * Note: The tables below are dependent on this value.
1859 #define LOAD_AVG_PERIOD 32
1860 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
1861 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
1863 /* Precomputed fixed inverse multiplies for multiplication by y^n */
1864 static const u32 runnable_avg_yN_inv[] = {
1865 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
1866 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
1867 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
1868 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
1869 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
1870 0x85aac367, 0x82cd8698,
1874 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
1875 * over-estimates when re-combining.
1877 static const u32 runnable_avg_yN_sum[] = {
1878 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
1879 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
1880 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
1885 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
1887 static __always_inline u64 decay_load(u64 val, u64 n)
1889 unsigned int local_n;
1893 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
1896 /* after bounds checking we can collapse to 32-bit */
1900 * As y^PERIOD = 1/2, we can combine
1901 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
1902 * With a look-up table which covers k^n (n<PERIOD)
1904 * To achieve constant time decay_load.
1906 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
1907 val >>= local_n / LOAD_AVG_PERIOD;
1908 local_n %= LOAD_AVG_PERIOD;
1911 val *= runnable_avg_yN_inv[local_n];
1912 /* We don't use SRR here since we always want to round down. */
1917 * For updates fully spanning n periods, the contribution to runnable
1918 * average will be: \Sum 1024*y^n
1920 * We can compute this reasonably efficiently by combining:
1921 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
1923 static u32 __compute_runnable_contrib(u64 n)
1927 if (likely(n <= LOAD_AVG_PERIOD))
1928 return runnable_avg_yN_sum[n];
1929 else if (unlikely(n >= LOAD_AVG_MAX_N))
1930 return LOAD_AVG_MAX;
1932 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
1934 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
1935 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
1937 n -= LOAD_AVG_PERIOD;
1938 } while (n > LOAD_AVG_PERIOD);
1940 contrib = decay_load(contrib, n);
1941 return contrib + runnable_avg_yN_sum[n];
1945 * We can represent the historical contribution to runnable average as the
1946 * coefficients of a geometric series. To do this we sub-divide our runnable
1947 * history into segments of approximately 1ms (1024us); label the segment that
1948 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
1950 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
1952 * (now) (~1ms ago) (~2ms ago)
1954 * Let u_i denote the fraction of p_i that the entity was runnable.
1956 * We then designate the fractions u_i as our co-efficients, yielding the
1957 * following representation of historical load:
1958 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
1960 * We choose y based on the with of a reasonably scheduling period, fixing:
1963 * This means that the contribution to load ~32ms ago (u_32) will be weighted
1964 * approximately half as much as the contribution to load within the last ms
1967 * When a period "rolls over" and we have new u_0`, multiplying the previous
1968 * sum again by y is sufficient to update:
1969 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1970 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1972 static __always_inline int __update_entity_runnable_avg(u64 now,
1973 struct sched_avg *sa,
1977 u32 runnable_contrib;
1978 int delta_w, decayed = 0;
1980 delta = now - sa->last_runnable_update;
1982 * This should only happen when time goes backwards, which it
1983 * unfortunately does during sched clock init when we swap over to TSC.
1985 if ((s64)delta < 0) {
1986 sa->last_runnable_update = now;
1991 * Use 1024ns as the unit of measurement since it's a reasonable
1992 * approximation of 1us and fast to compute.
1997 sa->last_runnable_update = now;
1999 /* delta_w is the amount already accumulated against our next period */
2000 delta_w = sa->runnable_avg_period % 1024;
2001 if (delta + delta_w >= 1024) {
2002 /* period roll-over */
2006 * Now that we know we're crossing a period boundary, figure
2007 * out how much from delta we need to complete the current
2008 * period and accrue it.
2010 delta_w = 1024 - delta_w;
2012 sa->runnable_avg_sum += delta_w;
2013 sa->runnable_avg_period += delta_w;
2017 /* Figure out how many additional periods this update spans */
2018 periods = delta / 1024;
2021 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
2023 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
2026 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2027 runnable_contrib = __compute_runnable_contrib(periods);
2029 sa->runnable_avg_sum += runnable_contrib;
2030 sa->runnable_avg_period += runnable_contrib;
2033 /* Remainder of delta accrued against u_0` */
2035 sa->runnable_avg_sum += delta;
2036 sa->runnable_avg_period += delta;
2041 /* Synchronize an entity's decay with its parenting cfs_rq.*/
2042 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
2044 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2045 u64 decays = atomic64_read(&cfs_rq->decay_counter);
2047 decays -= se->avg.decay_count;
2051 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
2052 se->avg.decay_count = 0;
2057 #ifdef CONFIG_FAIR_GROUP_SCHED
2058 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2061 struct task_group *tg = cfs_rq->tg;
2064 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
2065 tg_contrib -= cfs_rq->tg_load_contrib;
2067 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
2068 atomic_long_add(tg_contrib, &tg->load_avg);
2069 cfs_rq->tg_load_contrib += tg_contrib;
2074 * Aggregate cfs_rq runnable averages into an equivalent task_group
2075 * representation for computing load contributions.
2077 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2078 struct cfs_rq *cfs_rq)
2080 struct task_group *tg = cfs_rq->tg;
2083 /* The fraction of a cpu used by this cfs_rq */
2084 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
2085 sa->runnable_avg_period + 1);
2086 contrib -= cfs_rq->tg_runnable_contrib;
2088 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
2089 atomic_add(contrib, &tg->runnable_avg);
2090 cfs_rq->tg_runnable_contrib += contrib;
2094 static inline void __update_group_entity_contrib(struct sched_entity *se)
2096 struct cfs_rq *cfs_rq = group_cfs_rq(se);
2097 struct task_group *tg = cfs_rq->tg;
2102 contrib = cfs_rq->tg_load_contrib * tg->shares;
2103 se->avg.load_avg_contrib = div_u64(contrib,
2104 atomic_long_read(&tg->load_avg) + 1);
2107 * For group entities we need to compute a correction term in the case
2108 * that they are consuming <1 cpu so that we would contribute the same
2109 * load as a task of equal weight.
2111 * Explicitly co-ordinating this measurement would be expensive, but
2112 * fortunately the sum of each cpus contribution forms a usable
2113 * lower-bound on the true value.
2115 * Consider the aggregate of 2 contributions. Either they are disjoint
2116 * (and the sum represents true value) or they are disjoint and we are
2117 * understating by the aggregate of their overlap.
2119 * Extending this to N cpus, for a given overlap, the maximum amount we
2120 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
2121 * cpus that overlap for this interval and w_i is the interval width.
2123 * On a small machine; the first term is well-bounded which bounds the
2124 * total error since w_i is a subset of the period. Whereas on a
2125 * larger machine, while this first term can be larger, if w_i is the
2126 * of consequential size guaranteed to see n_i*w_i quickly converge to
2127 * our upper bound of 1-cpu.
2129 runnable_avg = atomic_read(&tg->runnable_avg);
2130 if (runnable_avg < NICE_0_LOAD) {
2131 se->avg.load_avg_contrib *= runnable_avg;
2132 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
2136 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
2137 int force_update) {}
2138 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
2139 struct cfs_rq *cfs_rq) {}
2140 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
2143 static inline void __update_task_entity_contrib(struct sched_entity *se)
2147 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
2148 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
2149 contrib /= (se->avg.runnable_avg_period + 1);
2150 se->avg.load_avg_contrib = scale_load(contrib);
2153 /* Compute the current contribution to load_avg by se, return any delta */
2154 static long __update_entity_load_avg_contrib(struct sched_entity *se)
2156 long old_contrib = se->avg.load_avg_contrib;
2158 if (entity_is_task(se)) {
2159 __update_task_entity_contrib(se);
2161 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
2162 __update_group_entity_contrib(se);
2165 return se->avg.load_avg_contrib - old_contrib;
2168 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
2171 if (likely(load_contrib < cfs_rq->blocked_load_avg))
2172 cfs_rq->blocked_load_avg -= load_contrib;
2174 cfs_rq->blocked_load_avg = 0;
2177 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
2179 /* Update a sched_entity's runnable average */
2180 static inline void update_entity_load_avg(struct sched_entity *se,
2183 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2188 * For a group entity we need to use their owned cfs_rq_clock_task() in
2189 * case they are the parent of a throttled hierarchy.
2191 if (entity_is_task(se))
2192 now = cfs_rq_clock_task(cfs_rq);
2194 now = cfs_rq_clock_task(group_cfs_rq(se));
2196 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
2199 contrib_delta = __update_entity_load_avg_contrib(se);
2205 cfs_rq->runnable_load_avg += contrib_delta;
2207 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
2211 * Decay the load contributed by all blocked children and account this so that
2212 * their contribution may appropriately discounted when they wake up.
2214 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
2216 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
2219 decays = now - cfs_rq->last_decay;
2220 if (!decays && !force_update)
2223 if (atomic_long_read(&cfs_rq->removed_load)) {
2224 unsigned long removed_load;
2225 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0);
2226 subtract_blocked_load_contrib(cfs_rq, removed_load);
2230 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
2232 atomic64_add(decays, &cfs_rq->decay_counter);
2233 cfs_rq->last_decay = now;
2236 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
2239 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
2241 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable);
2242 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
2245 /* Add the load generated by se into cfs_rq's child load-average */
2246 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2247 struct sched_entity *se,
2251 * We track migrations using entity decay_count <= 0, on a wake-up
2252 * migration we use a negative decay count to track the remote decays
2253 * accumulated while sleeping.
2255 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they
2256 * are seen by enqueue_entity_load_avg() as a migration with an already
2257 * constructed load_avg_contrib.
2259 if (unlikely(se->avg.decay_count <= 0)) {
2260 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq));
2261 if (se->avg.decay_count) {
2263 * In a wake-up migration we have to approximate the
2264 * time sleeping. This is because we can't synchronize
2265 * clock_task between the two cpus, and it is not
2266 * guaranteed to be read-safe. Instead, we can
2267 * approximate this using our carried decays, which are
2268 * explicitly atomically readable.
2270 se->avg.last_runnable_update -= (-se->avg.decay_count)
2272 update_entity_load_avg(se, 0);
2273 /* Indicate that we're now synchronized and on-rq */
2274 se->avg.decay_count = 0;
2279 * Task re-woke on same cpu (or else migrate_task_rq_fair()
2280 * would have made count negative); we must be careful to avoid
2281 * double-accounting blocked time after synchronizing decays.
2283 se->avg.last_runnable_update += __synchronize_entity_decay(se)
2287 /* migrated tasks did not contribute to our blocked load */
2289 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
2290 update_entity_load_avg(se, 0);
2293 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
2294 /* we force update consideration on load-balancer moves */
2295 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
2299 * Remove se's load from this cfs_rq child load-average, if the entity is
2300 * transitioning to a blocked state we track its projected decay using
2303 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2304 struct sched_entity *se,
2307 update_entity_load_avg(se, 1);
2308 /* we force update consideration on load-balancer moves */
2309 update_cfs_rq_blocked_load(cfs_rq, !sleep);
2311 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
2313 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
2314 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
2315 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
2319 * Update the rq's load with the elapsed running time before entering
2320 * idle. if the last scheduled task is not a CFS task, idle_enter will
2321 * be the only way to update the runnable statistic.
2323 void idle_enter_fair(struct rq *this_rq)
2325 update_rq_runnable_avg(this_rq, 1);
2329 * Update the rq's load with the elapsed idle time before a task is
2330 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will
2331 * be the only way to update the runnable statistic.
2333 void idle_exit_fair(struct rq *this_rq)
2335 update_rq_runnable_avg(this_rq, 0);
2339 static inline void update_entity_load_avg(struct sched_entity *se,
2340 int update_cfs_rq) {}
2341 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
2342 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
2343 struct sched_entity *se,
2345 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
2346 struct sched_entity *se,
2348 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
2349 int force_update) {}
2352 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
2354 #ifdef CONFIG_SCHEDSTATS
2355 struct task_struct *tsk = NULL;
2357 if (entity_is_task(se))
2360 if (se->statistics.sleep_start) {
2361 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
2366 if (unlikely(delta > se->statistics.sleep_max))
2367 se->statistics.sleep_max = delta;
2369 se->statistics.sleep_start = 0;
2370 se->statistics.sum_sleep_runtime += delta;
2373 account_scheduler_latency(tsk, delta >> 10, 1);
2374 trace_sched_stat_sleep(tsk, delta);
2377 if (se->statistics.block_start) {
2378 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
2383 if (unlikely(delta > se->statistics.block_max))
2384 se->statistics.block_max = delta;
2386 se->statistics.block_start = 0;
2387 se->statistics.sum_sleep_runtime += delta;
2390 if (tsk->in_iowait) {
2391 se->statistics.iowait_sum += delta;
2392 se->statistics.iowait_count++;
2393 trace_sched_stat_iowait(tsk, delta);
2396 trace_sched_stat_blocked(tsk, delta);
2399 * Blocking time is in units of nanosecs, so shift by
2400 * 20 to get a milliseconds-range estimation of the
2401 * amount of time that the task spent sleeping:
2403 if (unlikely(prof_on == SLEEP_PROFILING)) {
2404 profile_hits(SLEEP_PROFILING,
2405 (void *)get_wchan(tsk),
2408 account_scheduler_latency(tsk, delta >> 10, 0);
2414 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
2416 #ifdef CONFIG_SCHED_DEBUG
2417 s64 d = se->vruntime - cfs_rq->min_vruntime;
2422 if (d > 3*sysctl_sched_latency)
2423 schedstat_inc(cfs_rq, nr_spread_over);
2428 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
2430 u64 vruntime = cfs_rq->min_vruntime;
2433 * The 'current' period is already promised to the current tasks,
2434 * however the extra weight of the new task will slow them down a
2435 * little, place the new task so that it fits in the slot that
2436 * stays open at the end.
2438 if (initial && sched_feat(START_DEBIT))
2439 vruntime += sched_vslice(cfs_rq, se);
2441 /* sleeps up to a single latency don't count. */
2443 unsigned long thresh = sysctl_sched_latency;
2446 * Halve their sleep time's effect, to allow
2447 * for a gentler effect of sleepers:
2449 if (sched_feat(GENTLE_FAIR_SLEEPERS))
2455 /* ensure we never gain time by being placed backwards. */
2456 se->vruntime = max_vruntime(se->vruntime, vruntime);
2459 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
2462 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2465 * Update the normalized vruntime before updating min_vruntime
2466 * through calling update_curr().
2468 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
2469 se->vruntime += cfs_rq->min_vruntime;
2472 * Update run-time statistics of the 'current'.
2474 update_curr(cfs_rq);
2475 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
2476 account_entity_enqueue(cfs_rq, se);
2477 update_cfs_shares(cfs_rq);
2479 if (flags & ENQUEUE_WAKEUP) {
2480 place_entity(cfs_rq, se, 0);
2481 enqueue_sleeper(cfs_rq, se);
2484 update_stats_enqueue(cfs_rq, se);
2485 check_spread(cfs_rq, se);
2486 if (se != cfs_rq->curr)
2487 __enqueue_entity(cfs_rq, se);
2490 if (cfs_rq->nr_running == 1) {
2491 list_add_leaf_cfs_rq(cfs_rq);
2492 check_enqueue_throttle(cfs_rq);
2496 static void __clear_buddies_last(struct sched_entity *se)
2498 for_each_sched_entity(se) {
2499 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2500 if (cfs_rq->last == se)
2501 cfs_rq->last = NULL;
2507 static void __clear_buddies_next(struct sched_entity *se)
2509 for_each_sched_entity(se) {
2510 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2511 if (cfs_rq->next == se)
2512 cfs_rq->next = NULL;
2518 static void __clear_buddies_skip(struct sched_entity *se)
2520 for_each_sched_entity(se) {
2521 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2522 if (cfs_rq->skip == se)
2523 cfs_rq->skip = NULL;
2529 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
2531 if (cfs_rq->last == se)
2532 __clear_buddies_last(se);
2534 if (cfs_rq->next == se)
2535 __clear_buddies_next(se);
2537 if (cfs_rq->skip == se)
2538 __clear_buddies_skip(se);
2541 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2544 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
2547 * Update run-time statistics of the 'current'.
2549 update_curr(cfs_rq);
2550 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
2552 update_stats_dequeue(cfs_rq, se);
2553 if (flags & DEQUEUE_SLEEP) {
2554 #ifdef CONFIG_SCHEDSTATS
2555 if (entity_is_task(se)) {
2556 struct task_struct *tsk = task_of(se);
2558 if (tsk->state & TASK_INTERRUPTIBLE)
2559 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
2560 if (tsk->state & TASK_UNINTERRUPTIBLE)
2561 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
2566 clear_buddies(cfs_rq, se);
2568 if (se != cfs_rq->curr)
2569 __dequeue_entity(cfs_rq, se);
2571 account_entity_dequeue(cfs_rq, se);
2574 * Normalize the entity after updating the min_vruntime because the
2575 * update can refer to the ->curr item and we need to reflect this
2576 * movement in our normalized position.
2578 if (!(flags & DEQUEUE_SLEEP))
2579 se->vruntime -= cfs_rq->min_vruntime;
2581 /* return excess runtime on last dequeue */
2582 return_cfs_rq_runtime(cfs_rq);
2584 update_min_vruntime(cfs_rq);
2585 update_cfs_shares(cfs_rq);
2589 * Preempt the current task with a newly woken task if needed:
2592 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
2594 unsigned long ideal_runtime, delta_exec;
2595 struct sched_entity *se;
2598 ideal_runtime = sched_slice(cfs_rq, curr);
2599 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
2600 if (delta_exec > ideal_runtime) {
2601 resched_task(rq_of(cfs_rq)->curr);
2603 * The current task ran long enough, ensure it doesn't get
2604 * re-elected due to buddy favours.
2606 clear_buddies(cfs_rq, curr);
2611 * Ensure that a task that missed wakeup preemption by a
2612 * narrow margin doesn't have to wait for a full slice.
2613 * This also mitigates buddy induced latencies under load.
2615 if (delta_exec < sysctl_sched_min_granularity)
2618 se = __pick_first_entity(cfs_rq);
2619 delta = curr->vruntime - se->vruntime;
2624 if (delta > ideal_runtime)
2625 resched_task(rq_of(cfs_rq)->curr);
2629 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
2631 /* 'current' is not kept within the tree. */
2634 * Any task has to be enqueued before it get to execute on
2635 * a CPU. So account for the time it spent waiting on the
2638 update_stats_wait_end(cfs_rq, se);
2639 __dequeue_entity(cfs_rq, se);
2642 update_stats_curr_start(cfs_rq, se);
2644 #ifdef CONFIG_SCHEDSTATS
2646 * Track our maximum slice length, if the CPU's load is at
2647 * least twice that of our own weight (i.e. dont track it
2648 * when there are only lesser-weight tasks around):
2650 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
2651 se->statistics.slice_max = max(se->statistics.slice_max,
2652 se->sum_exec_runtime - se->prev_sum_exec_runtime);
2655 se->prev_sum_exec_runtime = se->sum_exec_runtime;
2659 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
2662 * Pick the next process, keeping these things in mind, in this order:
2663 * 1) keep things fair between processes/task groups
2664 * 2) pick the "next" process, since someone really wants that to run
2665 * 3) pick the "last" process, for cache locality
2666 * 4) do not run the "skip" process, if something else is available
2668 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
2670 struct sched_entity *se = __pick_first_entity(cfs_rq);
2671 struct sched_entity *left = se;
2674 * Avoid running the skip buddy, if running something else can
2675 * be done without getting too unfair.
2677 if (cfs_rq->skip == se) {
2678 struct sched_entity *second = __pick_next_entity(se);
2679 if (second && wakeup_preempt_entity(second, left) < 1)
2684 * Prefer last buddy, try to return the CPU to a preempted task.
2686 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
2690 * Someone really wants this to run. If it's not unfair, run it.
2692 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
2695 clear_buddies(cfs_rq, se);
2700 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
2702 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
2705 * If still on the runqueue then deactivate_task()
2706 * was not called and update_curr() has to be done:
2709 update_curr(cfs_rq);
2711 /* throttle cfs_rqs exceeding runtime */
2712 check_cfs_rq_runtime(cfs_rq);
2714 check_spread(cfs_rq, prev);
2716 update_stats_wait_start(cfs_rq, prev);
2717 /* Put 'current' back into the tree. */
2718 __enqueue_entity(cfs_rq, prev);
2719 /* in !on_rq case, update occurred at dequeue */
2720 update_entity_load_avg(prev, 1);
2722 cfs_rq->curr = NULL;
2726 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
2729 * Update run-time statistics of the 'current'.
2731 update_curr(cfs_rq);
2734 * Ensure that runnable average is periodically updated.
2736 update_entity_load_avg(curr, 1);
2737 update_cfs_rq_blocked_load(cfs_rq, 1);
2738 update_cfs_shares(cfs_rq);
2740 #ifdef CONFIG_SCHED_HRTICK
2742 * queued ticks are scheduled to match the slice, so don't bother
2743 * validating it and just reschedule.
2746 resched_task(rq_of(cfs_rq)->curr);
2750 * don't let the period tick interfere with the hrtick preemption
2752 if (!sched_feat(DOUBLE_TICK) &&
2753 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
2757 if (cfs_rq->nr_running > 1)
2758 check_preempt_tick(cfs_rq, curr);
2762 /**************************************************
2763 * CFS bandwidth control machinery
2766 #ifdef CONFIG_CFS_BANDWIDTH
2768 #ifdef HAVE_JUMP_LABEL
2769 static struct static_key __cfs_bandwidth_used;
2771 static inline bool cfs_bandwidth_used(void)
2773 return static_key_false(&__cfs_bandwidth_used);
2776 void account_cfs_bandwidth_used(int enabled, int was_enabled)
2778 /* only need to count groups transitioning between enabled/!enabled */
2779 if (enabled && !was_enabled)
2780 static_key_slow_inc(&__cfs_bandwidth_used);
2781 else if (!enabled && was_enabled)
2782 static_key_slow_dec(&__cfs_bandwidth_used);
2784 #else /* HAVE_JUMP_LABEL */
2785 static bool cfs_bandwidth_used(void)
2790 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
2791 #endif /* HAVE_JUMP_LABEL */
2794 * default period for cfs group bandwidth.
2795 * default: 0.1s, units: nanoseconds
2797 static inline u64 default_cfs_period(void)
2799 return 100000000ULL;
2802 static inline u64 sched_cfs_bandwidth_slice(void)
2804 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
2808 * Replenish runtime according to assigned quota and update expiration time.
2809 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
2810 * additional synchronization around rq->lock.
2812 * requires cfs_b->lock
2814 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
2818 if (cfs_b->quota == RUNTIME_INF)
2821 now = sched_clock_cpu(smp_processor_id());
2822 cfs_b->runtime = cfs_b->quota;
2823 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
2826 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2828 return &tg->cfs_bandwidth;
2831 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
2832 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2834 if (unlikely(cfs_rq->throttle_count))
2835 return cfs_rq->throttled_clock_task;
2837 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
2840 /* returns 0 on failure to allocate runtime */
2841 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2843 struct task_group *tg = cfs_rq->tg;
2844 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
2845 u64 amount = 0, min_amount, expires;
2847 /* note: this is a positive sum as runtime_remaining <= 0 */
2848 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
2850 raw_spin_lock(&cfs_b->lock);
2851 if (cfs_b->quota == RUNTIME_INF)
2852 amount = min_amount;
2855 * If the bandwidth pool has become inactive, then at least one
2856 * period must have elapsed since the last consumption.
2857 * Refresh the global state and ensure bandwidth timer becomes
2860 if (!cfs_b->timer_active) {
2861 __refill_cfs_bandwidth_runtime(cfs_b);
2862 __start_cfs_bandwidth(cfs_b);
2865 if (cfs_b->runtime > 0) {
2866 amount = min(cfs_b->runtime, min_amount);
2867 cfs_b->runtime -= amount;
2871 expires = cfs_b->runtime_expires;
2872 raw_spin_unlock(&cfs_b->lock);
2874 cfs_rq->runtime_remaining += amount;
2876 * we may have advanced our local expiration to account for allowed
2877 * spread between our sched_clock and the one on which runtime was
2880 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
2881 cfs_rq->runtime_expires = expires;
2883 return cfs_rq->runtime_remaining > 0;
2887 * Note: This depends on the synchronization provided by sched_clock and the
2888 * fact that rq->clock snapshots this value.
2890 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2894 /* if the deadline is ahead of our clock, nothing to do */
2895 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
2898 if (cfs_rq->runtime_remaining < 0)
2902 * If the local deadline has passed we have to consider the
2903 * possibility that our sched_clock is 'fast' and the global deadline
2904 * has not truly expired.
2906 * Fortunately we can check determine whether this the case by checking
2907 * whether the global deadline has advanced.
2910 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
2911 /* extend local deadline, drift is bounded above by 2 ticks */
2912 cfs_rq->runtime_expires += TICK_NSEC;
2914 /* global deadline is ahead, expiration has passed */
2915 cfs_rq->runtime_remaining = 0;
2919 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2920 unsigned long delta_exec)
2922 /* dock delta_exec before expiring quota (as it could span periods) */
2923 cfs_rq->runtime_remaining -= delta_exec;
2924 expire_cfs_rq_runtime(cfs_rq);
2926 if (likely(cfs_rq->runtime_remaining > 0))
2930 * if we're unable to extend our runtime we resched so that the active
2931 * hierarchy can be throttled
2933 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
2934 resched_task(rq_of(cfs_rq)->curr);
2937 static __always_inline
2938 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
2940 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
2943 __account_cfs_rq_runtime(cfs_rq, delta_exec);
2946 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2948 return cfs_bandwidth_used() && cfs_rq->throttled;
2951 /* check whether cfs_rq, or any parent, is throttled */
2952 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2954 return cfs_bandwidth_used() && cfs_rq->throttle_count;
2958 * Ensure that neither of the group entities corresponding to src_cpu or
2959 * dest_cpu are members of a throttled hierarchy when performing group
2960 * load-balance operations.
2962 static inline int throttled_lb_pair(struct task_group *tg,
2963 int src_cpu, int dest_cpu)
2965 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
2967 src_cfs_rq = tg->cfs_rq[src_cpu];
2968 dest_cfs_rq = tg->cfs_rq[dest_cpu];
2970 return throttled_hierarchy(src_cfs_rq) ||
2971 throttled_hierarchy(dest_cfs_rq);
2974 /* updated child weight may affect parent so we have to do this bottom up */
2975 static int tg_unthrottle_up(struct task_group *tg, void *data)
2977 struct rq *rq = data;
2978 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2980 cfs_rq->throttle_count--;
2982 if (!cfs_rq->throttle_count) {
2983 /* adjust cfs_rq_clock_task() */
2984 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
2985 cfs_rq->throttled_clock_task;
2992 static int tg_throttle_down(struct task_group *tg, void *data)
2994 struct rq *rq = data;
2995 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
2997 /* group is entering throttled state, stop time */
2998 if (!cfs_rq->throttle_count)
2999 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3000 cfs_rq->throttle_count++;
3005 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3007 struct rq *rq = rq_of(cfs_rq);
3008 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3009 struct sched_entity *se;
3010 long task_delta, dequeue = 1;
3012 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3014 /* freeze hierarchy runnable averages while throttled */
3016 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3019 task_delta = cfs_rq->h_nr_running;
3020 for_each_sched_entity(se) {
3021 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3022 /* throttled entity or throttle-on-deactivate */
3027 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3028 qcfs_rq->h_nr_running -= task_delta;
3030 if (qcfs_rq->load.weight)
3035 rq->nr_running -= task_delta;
3037 cfs_rq->throttled = 1;
3038 cfs_rq->throttled_clock = rq_clock(rq);
3039 raw_spin_lock(&cfs_b->lock);
3040 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3041 raw_spin_unlock(&cfs_b->lock);
3044 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3046 struct rq *rq = rq_of(cfs_rq);
3047 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3048 struct sched_entity *se;
3052 se = cfs_rq->tg->se[cpu_of(rq)];
3054 cfs_rq->throttled = 0;
3056 update_rq_clock(rq);
3058 raw_spin_lock(&cfs_b->lock);
3059 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3060 list_del_rcu(&cfs_rq->throttled_list);
3061 raw_spin_unlock(&cfs_b->lock);
3063 /* update hierarchical throttle state */
3064 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3066 if (!cfs_rq->load.weight)
3069 task_delta = cfs_rq->h_nr_running;
3070 for_each_sched_entity(se) {
3074 cfs_rq = cfs_rq_of(se);
3076 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3077 cfs_rq->h_nr_running += task_delta;
3079 if (cfs_rq_throttled(cfs_rq))
3084 rq->nr_running += task_delta;
3086 /* determine whether we need to wake up potentially idle cpu */
3087 if (rq->curr == rq->idle && rq->cfs.nr_running)
3088 resched_task(rq->curr);
3091 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
3092 u64 remaining, u64 expires)
3094 struct cfs_rq *cfs_rq;
3095 u64 runtime = remaining;
3098 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
3100 struct rq *rq = rq_of(cfs_rq);
3102 raw_spin_lock(&rq->lock);
3103 if (!cfs_rq_throttled(cfs_rq))
3106 runtime = -cfs_rq->runtime_remaining + 1;
3107 if (runtime > remaining)
3108 runtime = remaining;
3109 remaining -= runtime;
3111 cfs_rq->runtime_remaining += runtime;
3112 cfs_rq->runtime_expires = expires;
3114 /* we check whether we're throttled above */
3115 if (cfs_rq->runtime_remaining > 0)
3116 unthrottle_cfs_rq(cfs_rq);
3119 raw_spin_unlock(&rq->lock);
3130 * Responsible for refilling a task_group's bandwidth and unthrottling its
3131 * cfs_rqs as appropriate. If there has been no activity within the last
3132 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
3133 * used to track this state.
3135 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
3137 u64 runtime, runtime_expires;
3138 int idle = 1, throttled;
3140 raw_spin_lock(&cfs_b->lock);
3141 /* no need to continue the timer with no bandwidth constraint */
3142 if (cfs_b->quota == RUNTIME_INF)
3145 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3146 /* idle depends on !throttled (for the case of a large deficit) */
3147 idle = cfs_b->idle && !throttled;
3148 cfs_b->nr_periods += overrun;
3150 /* if we're going inactive then everything else can be deferred */
3154 __refill_cfs_bandwidth_runtime(cfs_b);
3157 /* mark as potentially idle for the upcoming period */
3162 /* account preceding periods in which throttling occurred */
3163 cfs_b->nr_throttled += overrun;
3166 * There are throttled entities so we must first use the new bandwidth
3167 * to unthrottle them before making it generally available. This
3168 * ensures that all existing debts will be paid before a new cfs_rq is
3171 runtime = cfs_b->runtime;
3172 runtime_expires = cfs_b->runtime_expires;
3176 * This check is repeated as we are holding onto the new bandwidth
3177 * while we unthrottle. This can potentially race with an unthrottled
3178 * group trying to acquire new bandwidth from the global pool.
3180 while (throttled && runtime > 0) {
3181 raw_spin_unlock(&cfs_b->lock);
3182 /* we can't nest cfs_b->lock while distributing bandwidth */
3183 runtime = distribute_cfs_runtime(cfs_b, runtime,
3185 raw_spin_lock(&cfs_b->lock);
3187 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
3190 /* return (any) remaining runtime */
3191 cfs_b->runtime = runtime;
3193 * While we are ensured activity in the period following an
3194 * unthrottle, this also covers the case in which the new bandwidth is
3195 * insufficient to cover the existing bandwidth deficit. (Forcing the
3196 * timer to remain active while there are any throttled entities.)
3201 cfs_b->timer_active = 0;
3202 raw_spin_unlock(&cfs_b->lock);
3207 /* a cfs_rq won't donate quota below this amount */
3208 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
3209 /* minimum remaining period time to redistribute slack quota */
3210 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
3211 /* how long we wait to gather additional slack before distributing */
3212 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
3214 /* are we near the end of the current quota period? */
3215 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
3217 struct hrtimer *refresh_timer = &cfs_b->period_timer;
3220 /* if the call-back is running a quota refresh is already occurring */
3221 if (hrtimer_callback_running(refresh_timer))
3224 /* is a quota refresh about to occur? */
3225 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
3226 if (remaining < min_expire)
3232 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
3234 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
3236 /* if there's a quota refresh soon don't bother with slack */
3237 if (runtime_refresh_within(cfs_b, min_left))
3240 start_bandwidth_timer(&cfs_b->slack_timer,
3241 ns_to_ktime(cfs_bandwidth_slack_period));
3244 /* we know any runtime found here is valid as update_curr() precedes return */
3245 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3247 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3248 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
3250 if (slack_runtime <= 0)
3253 raw_spin_lock(&cfs_b->lock);
3254 if (cfs_b->quota != RUNTIME_INF &&
3255 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
3256 cfs_b->runtime += slack_runtime;
3258 /* we are under rq->lock, defer unthrottling using a timer */
3259 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
3260 !list_empty(&cfs_b->throttled_cfs_rq))
3261 start_cfs_slack_bandwidth(cfs_b);
3263 raw_spin_unlock(&cfs_b->lock);
3265 /* even if it's not valid for return we don't want to try again */
3266 cfs_rq->runtime_remaining -= slack_runtime;
3269 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3271 if (!cfs_bandwidth_used())
3274 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
3277 __return_cfs_rq_runtime(cfs_rq);
3281 * This is done with a timer (instead of inline with bandwidth return) since
3282 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
3284 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
3286 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
3289 /* confirm we're still not at a refresh boundary */
3290 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
3293 raw_spin_lock(&cfs_b->lock);
3294 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
3295 runtime = cfs_b->runtime;
3298 expires = cfs_b->runtime_expires;
3299 raw_spin_unlock(&cfs_b->lock);
3304 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
3306 raw_spin_lock(&cfs_b->lock);
3307 if (expires == cfs_b->runtime_expires)
3308 cfs_b->runtime = runtime;
3309 raw_spin_unlock(&cfs_b->lock);
3313 * When a group wakes up we want to make sure that its quota is not already
3314 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
3315 * runtime as update_curr() throttling can not not trigger until it's on-rq.
3317 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
3319 if (!cfs_bandwidth_used())
3322 /* an active group must be handled by the update_curr()->put() path */
3323 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
3326 /* ensure the group is not already throttled */
3327 if (cfs_rq_throttled(cfs_rq))
3330 /* update runtime allocation */
3331 account_cfs_rq_runtime(cfs_rq, 0);
3332 if (cfs_rq->runtime_remaining <= 0)
3333 throttle_cfs_rq(cfs_rq);
3336 /* conditionally throttle active cfs_rq's from put_prev_entity() */
3337 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3339 if (!cfs_bandwidth_used())
3342 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
3346 * it's possible for a throttled entity to be forced into a running
3347 * state (e.g. set_curr_task), in this case we're finished.
3349 if (cfs_rq_throttled(cfs_rq))
3352 throttle_cfs_rq(cfs_rq);
3355 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
3357 struct cfs_bandwidth *cfs_b =
3358 container_of(timer, struct cfs_bandwidth, slack_timer);
3359 do_sched_cfs_slack_timer(cfs_b);
3361 return HRTIMER_NORESTART;
3364 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
3366 struct cfs_bandwidth *cfs_b =
3367 container_of(timer, struct cfs_bandwidth, period_timer);
3373 now = hrtimer_cb_get_time(timer);
3374 overrun = hrtimer_forward(timer, now, cfs_b->period);
3379 idle = do_sched_cfs_period_timer(cfs_b, overrun);
3382 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
3385 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3387 raw_spin_lock_init(&cfs_b->lock);
3389 cfs_b->quota = RUNTIME_INF;
3390 cfs_b->period = ns_to_ktime(default_cfs_period());
3392 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
3393 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3394 cfs_b->period_timer.function = sched_cfs_period_timer;
3395 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3396 cfs_b->slack_timer.function = sched_cfs_slack_timer;
3399 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3401 cfs_rq->runtime_enabled = 0;
3402 INIT_LIST_HEAD(&cfs_rq->throttled_list);
3405 /* requires cfs_b->lock, may release to reprogram timer */
3406 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3409 * The timer may be active because we're trying to set a new bandwidth
3410 * period or because we're racing with the tear-down path
3411 * (timer_active==0 becomes visible before the hrtimer call-back
3412 * terminates). In either case we ensure that it's re-programmed
3414 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
3415 raw_spin_unlock(&cfs_b->lock);
3416 /* ensure cfs_b->lock is available while we wait */
3417 hrtimer_cancel(&cfs_b->period_timer);
3419 raw_spin_lock(&cfs_b->lock);
3420 /* if someone else restarted the timer then we're done */
3421 if (cfs_b->timer_active)
3425 cfs_b->timer_active = 1;
3426 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
3429 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
3431 hrtimer_cancel(&cfs_b->period_timer);
3432 hrtimer_cancel(&cfs_b->slack_timer);
3435 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
3437 struct cfs_rq *cfs_rq;
3439 for_each_leaf_cfs_rq(rq, cfs_rq) {
3440 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3442 if (!cfs_rq->runtime_enabled)
3446 * clock_task is not advancing so we just need to make sure
3447 * there's some valid quota amount
3449 cfs_rq->runtime_remaining = cfs_b->quota;
3450 if (cfs_rq_throttled(cfs_rq))
3451 unthrottle_cfs_rq(cfs_rq);
3455 #else /* CONFIG_CFS_BANDWIDTH */
3456 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3458 return rq_clock_task(rq_of(cfs_rq));
3461 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
3462 unsigned long delta_exec) {}
3463 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3464 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
3465 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3467 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3472 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3477 static inline int throttled_lb_pair(struct task_group *tg,
3478 int src_cpu, int dest_cpu)
3483 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3485 #ifdef CONFIG_FAIR_GROUP_SCHED
3486 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
3489 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3493 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
3494 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
3496 #endif /* CONFIG_CFS_BANDWIDTH */
3498 /**************************************************
3499 * CFS operations on tasks:
3502 #ifdef CONFIG_SCHED_HRTICK
3503 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
3505 struct sched_entity *se = &p->se;
3506 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3508 WARN_ON(task_rq(p) != rq);
3510 if (cfs_rq->nr_running > 1) {
3511 u64 slice = sched_slice(cfs_rq, se);
3512 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
3513 s64 delta = slice - ran;
3522 * Don't schedule slices shorter than 10000ns, that just
3523 * doesn't make sense. Rely on vruntime for fairness.
3526 delta = max_t(s64, 10000LL, delta);
3528 hrtick_start(rq, delta);
3533 * called from enqueue/dequeue and updates the hrtick when the
3534 * current task is from our class and nr_running is low enough
3537 static void hrtick_update(struct rq *rq)
3539 struct task_struct *curr = rq->curr;
3541 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
3544 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
3545 hrtick_start_fair(rq, curr);
3547 #else /* !CONFIG_SCHED_HRTICK */
3549 hrtick_start_fair(struct rq *rq, struct task_struct *p)
3553 static inline void hrtick_update(struct rq *rq)
3559 * The enqueue_task method is called before nr_running is
3560 * increased. Here we update the fair scheduling stats and
3561 * then put the task into the rbtree:
3564 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3566 struct cfs_rq *cfs_rq;
3567 struct sched_entity *se = &p->se;
3569 for_each_sched_entity(se) {
3572 cfs_rq = cfs_rq_of(se);
3573 enqueue_entity(cfs_rq, se, flags);
3576 * end evaluation on encountering a throttled cfs_rq
3578 * note: in the case of encountering a throttled cfs_rq we will
3579 * post the final h_nr_running increment below.
3581 if (cfs_rq_throttled(cfs_rq))
3583 cfs_rq->h_nr_running++;
3585 flags = ENQUEUE_WAKEUP;
3588 for_each_sched_entity(se) {
3589 cfs_rq = cfs_rq_of(se);
3590 cfs_rq->h_nr_running++;
3592 if (cfs_rq_throttled(cfs_rq))
3595 update_cfs_shares(cfs_rq);
3596 update_entity_load_avg(se, 1);
3600 update_rq_runnable_avg(rq, rq->nr_running);
3606 static void set_next_buddy(struct sched_entity *se);
3609 * The dequeue_task method is called before nr_running is
3610 * decreased. We remove the task from the rbtree and
3611 * update the fair scheduling stats:
3613 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
3615 struct cfs_rq *cfs_rq;
3616 struct sched_entity *se = &p->se;
3617 int task_sleep = flags & DEQUEUE_SLEEP;
3619 for_each_sched_entity(se) {
3620 cfs_rq = cfs_rq_of(se);
3621 dequeue_entity(cfs_rq, se, flags);
3624 * end evaluation on encountering a throttled cfs_rq
3626 * note: in the case of encountering a throttled cfs_rq we will
3627 * post the final h_nr_running decrement below.
3629 if (cfs_rq_throttled(cfs_rq))
3631 cfs_rq->h_nr_running--;
3633 /* Don't dequeue parent if it has other entities besides us */
3634 if (cfs_rq->load.weight) {
3636 * Bias pick_next to pick a task from this cfs_rq, as
3637 * p is sleeping when it is within its sched_slice.
3639 if (task_sleep && parent_entity(se))
3640 set_next_buddy(parent_entity(se));
3642 /* avoid re-evaluating load for this entity */
3643 se = parent_entity(se);
3646 flags |= DEQUEUE_SLEEP;
3649 for_each_sched_entity(se) {
3650 cfs_rq = cfs_rq_of(se);
3651 cfs_rq->h_nr_running--;
3653 if (cfs_rq_throttled(cfs_rq))
3656 update_cfs_shares(cfs_rq);
3657 update_entity_load_avg(se, 1);
3662 update_rq_runnable_avg(rq, 1);
3668 /* Used instead of source_load when we know the type == 0 */
3669 static unsigned long weighted_cpuload(const int cpu)
3671 return cpu_rq(cpu)->cfs.runnable_load_avg;
3675 * Return a low guess at the load of a migration-source cpu weighted
3676 * according to the scheduling class and "nice" value.
3678 * We want to under-estimate the load of migration sources, to
3679 * balance conservatively.
3681 static unsigned long source_load(int cpu, int type)
3683 struct rq *rq = cpu_rq(cpu);
3684 unsigned long total = weighted_cpuload(cpu);
3686 if (type == 0 || !sched_feat(LB_BIAS))
3689 return min(rq->cpu_load[type-1], total);
3693 * Return a high guess at the load of a migration-target cpu weighted
3694 * according to the scheduling class and "nice" value.
3696 static unsigned long target_load(int cpu, int type)
3698 struct rq *rq = cpu_rq(cpu);
3699 unsigned long total = weighted_cpuload(cpu);
3701 if (type == 0 || !sched_feat(LB_BIAS))
3704 return max(rq->cpu_load[type-1], total);
3707 static unsigned long power_of(int cpu)
3709 return cpu_rq(cpu)->cpu_power;
3712 static unsigned long cpu_avg_load_per_task(int cpu)
3714 struct rq *rq = cpu_rq(cpu);
3715 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
3716 unsigned long load_avg = rq->cfs.runnable_load_avg;
3719 return load_avg / nr_running;
3724 static void record_wakee(struct task_struct *p)
3727 * Rough decay (wiping) for cost saving, don't worry
3728 * about the boundary, really active task won't care
3731 if (jiffies > current->wakee_flip_decay_ts + HZ) {
3732 current->wakee_flips = 0;
3733 current->wakee_flip_decay_ts = jiffies;
3736 if (current->last_wakee != p) {
3737 current->last_wakee = p;
3738 current->wakee_flips++;
3742 static void task_waking_fair(struct task_struct *p)
3744 struct sched_entity *se = &p->se;
3745 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3748 #ifndef CONFIG_64BIT
3749 u64 min_vruntime_copy;
3752 min_vruntime_copy = cfs_rq->min_vruntime_copy;
3754 min_vruntime = cfs_rq->min_vruntime;
3755 } while (min_vruntime != min_vruntime_copy);
3757 min_vruntime = cfs_rq->min_vruntime;
3760 se->vruntime -= min_vruntime;
3764 #ifdef CONFIG_FAIR_GROUP_SCHED
3766 * effective_load() calculates the load change as seen from the root_task_group
3768 * Adding load to a group doesn't make a group heavier, but can cause movement
3769 * of group shares between cpus. Assuming the shares were perfectly aligned one
3770 * can calculate the shift in shares.
3772 * Calculate the effective load difference if @wl is added (subtracted) to @tg
3773 * on this @cpu and results in a total addition (subtraction) of @wg to the
3774 * total group weight.
3776 * Given a runqueue weight distribution (rw_i) we can compute a shares
3777 * distribution (s_i) using:
3779 * s_i = rw_i / \Sum rw_j (1)
3781 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
3782 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
3783 * shares distribution (s_i):
3785 * rw_i = { 2, 4, 1, 0 }
3786 * s_i = { 2/7, 4/7, 1/7, 0 }
3788 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
3789 * task used to run on and the CPU the waker is running on), we need to
3790 * compute the effect of waking a task on either CPU and, in case of a sync
3791 * wakeup, compute the effect of the current task going to sleep.
3793 * So for a change of @wl to the local @cpu with an overall group weight change
3794 * of @wl we can compute the new shares distribution (s'_i) using:
3796 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
3798 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
3799 * differences in waking a task to CPU 0. The additional task changes the
3800 * weight and shares distributions like:
3802 * rw'_i = { 3, 4, 1, 0 }
3803 * s'_i = { 3/8, 4/8, 1/8, 0 }
3805 * We can then compute the difference in effective weight by using:
3807 * dw_i = S * (s'_i - s_i) (3)
3809 * Where 'S' is the group weight as seen by its parent.
3811 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
3812 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
3813 * 4/7) times the weight of the group.
3815 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3817 struct sched_entity *se = tg->se[cpu];
3819 if (!tg->parent || !wl) /* the trivial, non-cgroup case */
3822 for_each_sched_entity(se) {
3828 * W = @wg + \Sum rw_j
3830 W = wg + calc_tg_weight(tg, se->my_q);
3835 w = se->my_q->load.weight + wl;
3838 * wl = S * s'_i; see (2)
3841 wl = (w * tg->shares) / W;
3846 * Per the above, wl is the new se->load.weight value; since
3847 * those are clipped to [MIN_SHARES, ...) do so now. See
3848 * calc_cfs_shares().
3850 if (wl < MIN_SHARES)
3854 * wl = dw_i = S * (s'_i - s_i); see (3)
3856 wl -= se->load.weight;
3859 * Recursively apply this logic to all parent groups to compute
3860 * the final effective load change on the root group. Since
3861 * only the @tg group gets extra weight, all parent groups can
3862 * only redistribute existing shares. @wl is the shift in shares
3863 * resulting from this level per the above.
3872 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
3879 static int wake_wide(struct task_struct *p)
3881 int factor = this_cpu_read(sd_llc_size);
3884 * Yeah, it's the switching-frequency, could means many wakee or
3885 * rapidly switch, use factor here will just help to automatically
3886 * adjust the loose-degree, so bigger node will lead to more pull.
3888 if (p->wakee_flips > factor) {
3890 * wakee is somewhat hot, it needs certain amount of cpu
3891 * resource, so if waker is far more hot, prefer to leave
3894 if (current->wakee_flips > (factor * p->wakee_flips))
3901 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
3903 s64 this_load, load;
3904 int idx, this_cpu, prev_cpu;
3905 unsigned long tl_per_task;
3906 struct task_group *tg;
3907 unsigned long weight;
3911 * If we wake multiple tasks be careful to not bounce
3912 * ourselves around too much.
3918 this_cpu = smp_processor_id();
3919 prev_cpu = task_cpu(p);
3920 load = source_load(prev_cpu, idx);
3921 this_load = target_load(this_cpu, idx);
3924 * If sync wakeup then subtract the (maximum possible)
3925 * effect of the currently running task from the load
3926 * of the current CPU:
3929 tg = task_group(current);
3930 weight = current->se.load.weight;
3932 this_load += effective_load(tg, this_cpu, -weight, -weight);
3933 load += effective_load(tg, prev_cpu, 0, -weight);
3937 weight = p->se.load.weight;
3940 * In low-load situations, where prev_cpu is idle and this_cpu is idle
3941 * due to the sync cause above having dropped this_load to 0, we'll
3942 * always have an imbalance, but there's really nothing you can do
3943 * about that, so that's good too.
3945 * Otherwise check if either cpus are near enough in load to allow this
3946 * task to be woken on this_cpu.
3948 if (this_load > 0) {
3949 s64 this_eff_load, prev_eff_load;
3951 this_eff_load = 100;
3952 this_eff_load *= power_of(prev_cpu);
3953 this_eff_load *= this_load +
3954 effective_load(tg, this_cpu, weight, weight);
3956 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
3957 prev_eff_load *= power_of(this_cpu);
3958 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
3960 balanced = this_eff_load <= prev_eff_load;
3965 * If the currently running task will sleep within
3966 * a reasonable amount of time then attract this newly
3969 if (sync && balanced)
3972 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
3973 tl_per_task = cpu_avg_load_per_task(this_cpu);
3976 (this_load <= load &&
3977 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
3979 * This domain has SD_WAKE_AFFINE and
3980 * p is cache cold in this domain, and
3981 * there is no bad imbalance.
3983 schedstat_inc(sd, ttwu_move_affine);
3984 schedstat_inc(p, se.statistics.nr_wakeups_affine);
3992 * find_idlest_group finds and returns the least busy CPU group within the
3995 static struct sched_group *
3996 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
3997 int this_cpu, int load_idx)
3999 struct sched_group *idlest = NULL, *group = sd->groups;
4000 unsigned long min_load = ULONG_MAX, this_load = 0;
4001 int imbalance = 100 + (sd->imbalance_pct-100)/2;
4004 unsigned long load, avg_load;
4008 /* Skip over this group if it has no CPUs allowed */
4009 if (!cpumask_intersects(sched_group_cpus(group),
4010 tsk_cpus_allowed(p)))
4013 local_group = cpumask_test_cpu(this_cpu,
4014 sched_group_cpus(group));
4016 /* Tally up the load of all CPUs in the group */
4019 for_each_cpu(i, sched_group_cpus(group)) {
4020 /* Bias balancing toward cpus of our domain */
4022 load = source_load(i, load_idx);
4024 load = target_load(i, load_idx);
4029 /* Adjust by relative CPU power of the group */
4030 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
4033 this_load = avg_load;
4034 } else if (avg_load < min_load) {
4035 min_load = avg_load;
4038 } while (group = group->next, group != sd->groups);
4040 if (!idlest || 100*this_load < imbalance*min_load)
4046 * find_idlest_cpu - find the idlest cpu among the cpus in group.
4049 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
4051 unsigned long load, min_load = ULONG_MAX;
4055 /* Traverse only the allowed CPUs */
4056 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
4057 load = weighted_cpuload(i);
4059 if (load < min_load || (load == min_load && i == this_cpu)) {
4069 * Try and locate an idle CPU in the sched_domain.
4071 static int select_idle_sibling(struct task_struct *p, int target)
4073 struct sched_domain *sd;
4074 struct sched_group *sg;
4075 int i = task_cpu(p);
4077 if (idle_cpu(target))
4081 * If the prevous cpu is cache affine and idle, don't be stupid.
4083 if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
4087 * Otherwise, iterate the domains and find an elegible idle cpu.
4089 sd = rcu_dereference(per_cpu(sd_llc, target));
4090 for_each_lower_domain(sd) {
4093 if (!cpumask_intersects(sched_group_cpus(sg),
4094 tsk_cpus_allowed(p)))
4097 for_each_cpu(i, sched_group_cpus(sg)) {
4098 if (i == target || !idle_cpu(i))
4102 target = cpumask_first_and(sched_group_cpus(sg),
4103 tsk_cpus_allowed(p));
4107 } while (sg != sd->groups);
4114 * sched_balance_self: balance the current task (running on cpu) in domains
4115 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
4118 * Balance, ie. select the least loaded group.
4120 * Returns the target CPU number, or the same CPU if no balancing is needed.
4122 * preempt must be disabled.
4125 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
4127 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
4128 int cpu = smp_processor_id();
4130 int want_affine = 0;
4131 int sync = wake_flags & WF_SYNC;
4133 if (p->nr_cpus_allowed == 1)
4136 if (sd_flag & SD_BALANCE_WAKE) {
4137 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
4143 for_each_domain(cpu, tmp) {
4144 if (!(tmp->flags & SD_LOAD_BALANCE))
4148 * If both cpu and prev_cpu are part of this domain,
4149 * cpu is a valid SD_WAKE_AFFINE target.
4151 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
4152 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
4157 if (tmp->flags & sd_flag)
4162 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
4165 new_cpu = select_idle_sibling(p, prev_cpu);
4170 int load_idx = sd->forkexec_idx;
4171 struct sched_group *group;
4174 if (!(sd->flags & sd_flag)) {
4179 if (sd_flag & SD_BALANCE_WAKE)
4180 load_idx = sd->wake_idx;
4182 group = find_idlest_group(sd, p, cpu, load_idx);
4188 new_cpu = find_idlest_cpu(group, p, cpu);
4189 if (new_cpu == -1 || new_cpu == cpu) {
4190 /* Now try balancing at a lower domain level of cpu */
4195 /* Now try balancing at a lower domain level of new_cpu */
4197 weight = sd->span_weight;
4199 for_each_domain(cpu, tmp) {
4200 if (weight <= tmp->span_weight)
4202 if (tmp->flags & sd_flag)
4205 /* while loop will break here if sd == NULL */
4214 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
4215 * cfs_rq_of(p) references at time of call are still valid and identify the
4216 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
4217 * other assumptions, including the state of rq->lock, should be made.
4220 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
4222 struct sched_entity *se = &p->se;
4223 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4226 * Load tracking: accumulate removed load so that it can be processed
4227 * when we next update owning cfs_rq under rq->lock. Tasks contribute
4228 * to blocked load iff they have a positive decay-count. It can never
4229 * be negative here since on-rq tasks have decay-count == 0.
4231 if (se->avg.decay_count) {
4232 se->avg.decay_count = -__synchronize_entity_decay(se);
4233 atomic_long_add(se->avg.load_avg_contrib,
4234 &cfs_rq->removed_load);
4237 #endif /* CONFIG_SMP */
4239 static unsigned long
4240 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
4242 unsigned long gran = sysctl_sched_wakeup_granularity;
4245 * Since its curr running now, convert the gran from real-time
4246 * to virtual-time in his units.
4248 * By using 'se' instead of 'curr' we penalize light tasks, so
4249 * they get preempted easier. That is, if 'se' < 'curr' then
4250 * the resulting gran will be larger, therefore penalizing the
4251 * lighter, if otoh 'se' > 'curr' then the resulting gran will
4252 * be smaller, again penalizing the lighter task.
4254 * This is especially important for buddies when the leftmost
4255 * task is higher priority than the buddy.
4257 return calc_delta_fair(gran, se);
4261 * Should 'se' preempt 'curr'.
4275 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
4277 s64 gran, vdiff = curr->vruntime - se->vruntime;
4282 gran = wakeup_gran(curr, se);
4289 static void set_last_buddy(struct sched_entity *se)
4291 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4294 for_each_sched_entity(se)
4295 cfs_rq_of(se)->last = se;
4298 static void set_next_buddy(struct sched_entity *se)
4300 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
4303 for_each_sched_entity(se)
4304 cfs_rq_of(se)->next = se;
4307 static void set_skip_buddy(struct sched_entity *se)
4309 for_each_sched_entity(se)
4310 cfs_rq_of(se)->skip = se;
4314 * Preempt the current task with a newly woken task if needed:
4316 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
4318 struct task_struct *curr = rq->curr;
4319 struct sched_entity *se = &curr->se, *pse = &p->se;
4320 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4321 int scale = cfs_rq->nr_running >= sched_nr_latency;
4322 int next_buddy_marked = 0;
4324 if (unlikely(se == pse))
4328 * This is possible from callers such as move_task(), in which we
4329 * unconditionally check_prempt_curr() after an enqueue (which may have
4330 * lead to a throttle). This both saves work and prevents false
4331 * next-buddy nomination below.
4333 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
4336 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
4337 set_next_buddy(pse);
4338 next_buddy_marked = 1;
4342 * We can come here with TIF_NEED_RESCHED already set from new task
4345 * Note: this also catches the edge-case of curr being in a throttled
4346 * group (e.g. via set_curr_task), since update_curr() (in the
4347 * enqueue of curr) will have resulted in resched being set. This
4348 * prevents us from potentially nominating it as a false LAST_BUDDY
4351 if (test_tsk_need_resched(curr))
4354 /* Idle tasks are by definition preempted by non-idle tasks. */
4355 if (unlikely(curr->policy == SCHED_IDLE) &&
4356 likely(p->policy != SCHED_IDLE))
4360 * Batch and idle tasks do not preempt non-idle tasks (their preemption
4361 * is driven by the tick):
4363 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
4366 find_matching_se(&se, &pse);
4367 update_curr(cfs_rq_of(se));
4369 if (wakeup_preempt_entity(se, pse) == 1) {
4371 * Bias pick_next to pick the sched entity that is
4372 * triggering this preemption.
4374 if (!next_buddy_marked)
4375 set_next_buddy(pse);
4384 * Only set the backward buddy when the current task is still
4385 * on the rq. This can happen when a wakeup gets interleaved
4386 * with schedule on the ->pre_schedule() or idle_balance()
4387 * point, either of which can * drop the rq lock.
4389 * Also, during early boot the idle thread is in the fair class,
4390 * for obvious reasons its a bad idea to schedule back to it.
4392 if (unlikely(!se->on_rq || curr == rq->idle))
4395 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
4399 static struct task_struct *pick_next_task_fair(struct rq *rq)
4401 struct task_struct *p;
4402 struct cfs_rq *cfs_rq = &rq->cfs;
4403 struct sched_entity *se;
4405 if (!cfs_rq->nr_running)
4409 se = pick_next_entity(cfs_rq);
4410 set_next_entity(cfs_rq, se);
4411 cfs_rq = group_cfs_rq(se);
4415 if (hrtick_enabled(rq))
4416 hrtick_start_fair(rq, p);
4422 * Account for a descheduled task:
4424 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
4426 struct sched_entity *se = &prev->se;
4427 struct cfs_rq *cfs_rq;
4429 for_each_sched_entity(se) {
4430 cfs_rq = cfs_rq_of(se);
4431 put_prev_entity(cfs_rq, se);
4436 * sched_yield() is very simple
4438 * The magic of dealing with the ->skip buddy is in pick_next_entity.
4440 static void yield_task_fair(struct rq *rq)
4442 struct task_struct *curr = rq->curr;
4443 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
4444 struct sched_entity *se = &curr->se;
4447 * Are we the only task in the tree?
4449 if (unlikely(rq->nr_running == 1))
4452 clear_buddies(cfs_rq, se);
4454 if (curr->policy != SCHED_BATCH) {
4455 update_rq_clock(rq);
4457 * Update run-time statistics of the 'current'.
4459 update_curr(cfs_rq);
4461 * Tell update_rq_clock() that we've just updated,
4462 * so we don't do microscopic update in schedule()
4463 * and double the fastpath cost.
4465 rq->skip_clock_update = 1;
4471 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
4473 struct sched_entity *se = &p->se;
4475 /* throttled hierarchies are not runnable */
4476 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
4479 /* Tell the scheduler that we'd really like pse to run next. */
4482 yield_task_fair(rq);
4488 /**************************************************
4489 * Fair scheduling class load-balancing methods.
4493 * The purpose of load-balancing is to achieve the same basic fairness the
4494 * per-cpu scheduler provides, namely provide a proportional amount of compute
4495 * time to each task. This is expressed in the following equation:
4497 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
4499 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
4500 * W_i,0 is defined as:
4502 * W_i,0 = \Sum_j w_i,j (2)
4504 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
4505 * is derived from the nice value as per prio_to_weight[].
4507 * The weight average is an exponential decay average of the instantaneous
4510 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
4512 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
4513 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
4514 * can also include other factors [XXX].
4516 * To achieve this balance we define a measure of imbalance which follows
4517 * directly from (1):
4519 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
4521 * We them move tasks around to minimize the imbalance. In the continuous
4522 * function space it is obvious this converges, in the discrete case we get
4523 * a few fun cases generally called infeasible weight scenarios.
4526 * - infeasible weights;
4527 * - local vs global optima in the discrete case. ]
4532 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
4533 * for all i,j solution, we create a tree of cpus that follows the hardware
4534 * topology where each level pairs two lower groups (or better). This results
4535 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
4536 * tree to only the first of the previous level and we decrease the frequency
4537 * of load-balance at each level inv. proportional to the number of cpus in
4543 * \Sum { --- * --- * 2^i } = O(n) (5)
4545 * `- size of each group
4546 * | | `- number of cpus doing load-balance
4548 * `- sum over all levels
4550 * Coupled with a limit on how many tasks we can migrate every balance pass,
4551 * this makes (5) the runtime complexity of the balancer.
4553 * An important property here is that each CPU is still (indirectly) connected
4554 * to every other cpu in at most O(log n) steps:
4556 * The adjacency matrix of the resulting graph is given by:
4559 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
4562 * And you'll find that:
4564 * A^(log_2 n)_i,j != 0 for all i,j (7)
4566 * Showing there's indeed a path between every cpu in at most O(log n) steps.
4567 * The task movement gives a factor of O(m), giving a convergence complexity
4570 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
4575 * In order to avoid CPUs going idle while there's still work to do, new idle
4576 * balancing is more aggressive and has the newly idle cpu iterate up the domain
4577 * tree itself instead of relying on other CPUs to bring it work.
4579 * This adds some complexity to both (5) and (8) but it reduces the total idle
4587 * Cgroups make a horror show out of (2), instead of a simple sum we get:
4590 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
4595 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
4597 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
4599 * The big problem is S_k, its a global sum needed to compute a local (W_i)
4602 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
4603 * rewrite all of this once again.]
4606 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4608 #define LBF_ALL_PINNED 0x01
4609 #define LBF_NEED_BREAK 0x02
4610 #define LBF_DST_PINNED 0x04
4611 #define LBF_SOME_PINNED 0x08
4614 struct sched_domain *sd;
4622 struct cpumask *dst_grpmask;
4624 enum cpu_idle_type idle;
4626 /* The set of CPUs under consideration for load-balancing */
4627 struct cpumask *cpus;
4632 unsigned int loop_break;
4633 unsigned int loop_max;
4637 * move_task - move a task from one runqueue to another runqueue.
4638 * Both runqueues must be locked.
4640 static void move_task(struct task_struct *p, struct lb_env *env)
4642 deactivate_task(env->src_rq, p, 0);
4643 set_task_cpu(p, env->dst_cpu);
4644 activate_task(env->dst_rq, p, 0);
4645 check_preempt_curr(env->dst_rq, p, 0);
4646 #ifdef CONFIG_NUMA_BALANCING
4647 if (p->numa_preferred_nid != -1) {
4648 int src_nid = cpu_to_node(env->src_cpu);
4649 int dst_nid = cpu_to_node(env->dst_cpu);
4652 * If the load balancer has moved the task then limit
4653 * migrations from taking place in the short term in
4654 * case this is a short-lived migration.
4656 if (src_nid != dst_nid && dst_nid != p->numa_preferred_nid)
4657 p->numa_migrate_seq = 0;
4663 * Is this task likely cache-hot:
4666 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
4670 if (p->sched_class != &fair_sched_class)
4673 if (unlikely(p->policy == SCHED_IDLE))
4677 * Buddy candidates are cache hot:
4679 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
4680 (&p->se == cfs_rq_of(&p->se)->next ||
4681 &p->se == cfs_rq_of(&p->se)->last))
4684 if (sysctl_sched_migration_cost == -1)
4686 if (sysctl_sched_migration_cost == 0)
4689 delta = now - p->se.exec_start;
4691 return delta < (s64)sysctl_sched_migration_cost;
4694 #ifdef CONFIG_NUMA_BALANCING
4695 /* Returns true if the destination node has incurred more faults */
4696 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env)
4698 int src_nid, dst_nid;
4700 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults ||
4701 !(env->sd->flags & SD_NUMA)) {
4705 src_nid = cpu_to_node(env->src_cpu);
4706 dst_nid = cpu_to_node(env->dst_cpu);
4708 if (src_nid == dst_nid)
4711 /* Always encourage migration to the preferred node. */
4712 if (dst_nid == p->numa_preferred_nid)
4715 /* If both task and group weight improve, this move is a winner. */
4716 if (task_weight(p, dst_nid) > task_weight(p, src_nid) &&
4717 group_weight(p, dst_nid) > group_weight(p, src_nid))
4724 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
4726 int src_nid, dst_nid;
4728 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER))
4731 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
4734 src_nid = cpu_to_node(env->src_cpu);
4735 dst_nid = cpu_to_node(env->dst_cpu);
4737 if (src_nid == dst_nid)
4740 /* Migrating away from the preferred node is always bad. */
4741 if (src_nid == p->numa_preferred_nid)
4744 /* If either task or group weight get worse, don't do it. */
4745 if (task_weight(p, dst_nid) < task_weight(p, src_nid) ||
4746 group_weight(p, dst_nid) < group_weight(p, src_nid))
4753 static inline bool migrate_improves_locality(struct task_struct *p,
4759 static inline bool migrate_degrades_locality(struct task_struct *p,
4767 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
4770 int can_migrate_task(struct task_struct *p, struct lb_env *env)
4772 int tsk_cache_hot = 0;
4774 * We do not migrate tasks that are:
4775 * 1) throttled_lb_pair, or
4776 * 2) cannot be migrated to this CPU due to cpus_allowed, or
4777 * 3) running (obviously), or
4778 * 4) are cache-hot on their current CPU.
4780 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
4783 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
4786 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
4788 env->flags |= LBF_SOME_PINNED;
4791 * Remember if this task can be migrated to any other cpu in
4792 * our sched_group. We may want to revisit it if we couldn't
4793 * meet load balance goals by pulling other tasks on src_cpu.
4795 * Also avoid computing new_dst_cpu if we have already computed
4796 * one in current iteration.
4798 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
4801 /* Prevent to re-select dst_cpu via env's cpus */
4802 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
4803 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
4804 env->flags |= LBF_DST_PINNED;
4805 env->new_dst_cpu = cpu;
4813 /* Record that we found atleast one task that could run on dst_cpu */
4814 env->flags &= ~LBF_ALL_PINNED;
4816 if (task_running(env->src_rq, p)) {
4817 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
4822 * Aggressive migration if:
4823 * 1) destination numa is preferred
4824 * 2) task is cache cold, or
4825 * 3) too many balance attempts have failed.
4827 tsk_cache_hot = task_hot(p, rq_clock_task(env->src_rq), env->sd);
4829 tsk_cache_hot = migrate_degrades_locality(p, env);
4831 if (migrate_improves_locality(p, env)) {
4832 #ifdef CONFIG_SCHEDSTATS
4833 if (tsk_cache_hot) {
4834 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4835 schedstat_inc(p, se.statistics.nr_forced_migrations);
4841 if (!tsk_cache_hot ||
4842 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
4844 if (tsk_cache_hot) {
4845 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
4846 schedstat_inc(p, se.statistics.nr_forced_migrations);
4852 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
4857 * move_one_task tries to move exactly one task from busiest to this_rq, as
4858 * part of active balancing operations within "domain".
4859 * Returns 1 if successful and 0 otherwise.
4861 * Called with both runqueues locked.
4863 static int move_one_task(struct lb_env *env)
4865 struct task_struct *p, *n;
4867 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
4868 if (!can_migrate_task(p, env))
4873 * Right now, this is only the second place move_task()
4874 * is called, so we can safely collect move_task()
4875 * stats here rather than inside move_task().
4877 schedstat_inc(env->sd, lb_gained[env->idle]);
4883 static const unsigned int sched_nr_migrate_break = 32;
4886 * move_tasks tries to move up to imbalance weighted load from busiest to
4887 * this_rq, as part of a balancing operation within domain "sd".
4888 * Returns 1 if successful and 0 otherwise.
4890 * Called with both runqueues locked.
4892 static int move_tasks(struct lb_env *env)
4894 struct list_head *tasks = &env->src_rq->cfs_tasks;
4895 struct task_struct *p;
4899 if (env->imbalance <= 0)
4902 while (!list_empty(tasks)) {
4903 p = list_first_entry(tasks, struct task_struct, se.group_node);
4906 /* We've more or less seen every task there is, call it quits */
4907 if (env->loop > env->loop_max)
4910 /* take a breather every nr_migrate tasks */
4911 if (env->loop > env->loop_break) {
4912 env->loop_break += sched_nr_migrate_break;
4913 env->flags |= LBF_NEED_BREAK;
4917 if (!can_migrate_task(p, env))
4920 load = task_h_load(p);
4922 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
4925 if ((load / 2) > env->imbalance)
4930 env->imbalance -= load;
4932 #ifdef CONFIG_PREEMPT
4934 * NEWIDLE balancing is a source of latency, so preemptible
4935 * kernels will stop after the first task is pulled to minimize
4936 * the critical section.
4938 if (env->idle == CPU_NEWLY_IDLE)
4943 * We only want to steal up to the prescribed amount of
4946 if (env->imbalance <= 0)
4951 list_move_tail(&p->se.group_node, tasks);
4955 * Right now, this is one of only two places move_task() is called,
4956 * so we can safely collect move_task() stats here rather than
4957 * inside move_task().
4959 schedstat_add(env->sd, lb_gained[env->idle], pulled);
4964 #ifdef CONFIG_FAIR_GROUP_SCHED
4966 * update tg->load_weight by folding this cpu's load_avg
4968 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
4970 struct sched_entity *se = tg->se[cpu];
4971 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
4973 /* throttled entities do not contribute to load */
4974 if (throttled_hierarchy(cfs_rq))
4977 update_cfs_rq_blocked_load(cfs_rq, 1);
4980 update_entity_load_avg(se, 1);
4982 * We pivot on our runnable average having decayed to zero for
4983 * list removal. This generally implies that all our children
4984 * have also been removed (modulo rounding error or bandwidth
4985 * control); however, such cases are rare and we can fix these
4988 * TODO: fix up out-of-order children on enqueue.
4990 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
4991 list_del_leaf_cfs_rq(cfs_rq);
4993 struct rq *rq = rq_of(cfs_rq);
4994 update_rq_runnable_avg(rq, rq->nr_running);
4998 static void update_blocked_averages(int cpu)
5000 struct rq *rq = cpu_rq(cpu);
5001 struct cfs_rq *cfs_rq;
5002 unsigned long flags;
5004 raw_spin_lock_irqsave(&rq->lock, flags);
5005 update_rq_clock(rq);
5007 * Iterates the task_group tree in a bottom up fashion, see
5008 * list_add_leaf_cfs_rq() for details.
5010 for_each_leaf_cfs_rq(rq, cfs_rq) {
5012 * Note: We may want to consider periodically releasing
5013 * rq->lock about these updates so that creating many task
5014 * groups does not result in continually extending hold time.
5016 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
5019 raw_spin_unlock_irqrestore(&rq->lock, flags);
5023 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
5024 * This needs to be done in a top-down fashion because the load of a child
5025 * group is a fraction of its parents load.
5027 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
5029 struct rq *rq = rq_of(cfs_rq);
5030 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
5031 unsigned long now = jiffies;
5034 if (cfs_rq->last_h_load_update == now)
5037 cfs_rq->h_load_next = NULL;
5038 for_each_sched_entity(se) {
5039 cfs_rq = cfs_rq_of(se);
5040 cfs_rq->h_load_next = se;
5041 if (cfs_rq->last_h_load_update == now)
5046 cfs_rq->h_load = cfs_rq->runnable_load_avg;
5047 cfs_rq->last_h_load_update = now;
5050 while ((se = cfs_rq->h_load_next) != NULL) {
5051 load = cfs_rq->h_load;
5052 load = div64_ul(load * se->avg.load_avg_contrib,
5053 cfs_rq->runnable_load_avg + 1);
5054 cfs_rq = group_cfs_rq(se);
5055 cfs_rq->h_load = load;
5056 cfs_rq->last_h_load_update = now;
5060 static unsigned long task_h_load(struct task_struct *p)
5062 struct cfs_rq *cfs_rq = task_cfs_rq(p);
5064 update_cfs_rq_h_load(cfs_rq);
5065 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load,
5066 cfs_rq->runnable_load_avg + 1);
5069 static inline void update_blocked_averages(int cpu)
5073 static unsigned long task_h_load(struct task_struct *p)
5075 return p->se.avg.load_avg_contrib;
5079 /********** Helpers for find_busiest_group ************************/
5081 * sg_lb_stats - stats of a sched_group required for load_balancing
5083 struct sg_lb_stats {
5084 unsigned long avg_load; /*Avg load across the CPUs of the group */
5085 unsigned long group_load; /* Total load over the CPUs of the group */
5086 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
5087 unsigned long load_per_task;
5088 unsigned long group_power;
5089 unsigned int sum_nr_running; /* Nr tasks running in the group */
5090 unsigned int group_capacity;
5091 unsigned int idle_cpus;
5092 unsigned int group_weight;
5093 int group_imb; /* Is there an imbalance in the group ? */
5094 int group_has_capacity; /* Is there extra capacity in the group? */
5098 * sd_lb_stats - Structure to store the statistics of a sched_domain
5099 * during load balancing.
5101 struct sd_lb_stats {
5102 struct sched_group *busiest; /* Busiest group in this sd */
5103 struct sched_group *local; /* Local group in this sd */
5104 unsigned long total_load; /* Total load of all groups in sd */
5105 unsigned long total_pwr; /* Total power of all groups in sd */
5106 unsigned long avg_load; /* Average load across all groups in sd */
5108 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
5109 struct sg_lb_stats local_stat; /* Statistics of the local group */
5112 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
5115 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
5116 * local_stat because update_sg_lb_stats() does a full clear/assignment.
5117 * We must however clear busiest_stat::avg_load because
5118 * update_sd_pick_busiest() reads this before assignment.
5120 *sds = (struct sd_lb_stats){
5132 * get_sd_load_idx - Obtain the load index for a given sched domain.
5133 * @sd: The sched_domain whose load_idx is to be obtained.
5134 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
5136 * Return: The load index.
5138 static inline int get_sd_load_idx(struct sched_domain *sd,
5139 enum cpu_idle_type idle)
5145 load_idx = sd->busy_idx;
5148 case CPU_NEWLY_IDLE:
5149 load_idx = sd->newidle_idx;
5152 load_idx = sd->idle_idx;
5159 static unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
5161 return SCHED_POWER_SCALE;
5164 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
5166 return default_scale_freq_power(sd, cpu);
5169 static unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
5171 unsigned long weight = sd->span_weight;
5172 unsigned long smt_gain = sd->smt_gain;
5179 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
5181 return default_scale_smt_power(sd, cpu);
5184 static unsigned long scale_rt_power(int cpu)
5186 struct rq *rq = cpu_rq(cpu);
5187 u64 total, available, age_stamp, avg;
5190 * Since we're reading these variables without serialization make sure
5191 * we read them once before doing sanity checks on them.
5193 age_stamp = ACCESS_ONCE(rq->age_stamp);
5194 avg = ACCESS_ONCE(rq->rt_avg);
5196 total = sched_avg_period() + (rq_clock(rq) - age_stamp);
5198 if (unlikely(total < avg)) {
5199 /* Ensures that power won't end up being negative */
5202 available = total - avg;
5205 if (unlikely((s64)total < SCHED_POWER_SCALE))
5206 total = SCHED_POWER_SCALE;
5208 total >>= SCHED_POWER_SHIFT;
5210 return div_u64(available, total);
5213 static void update_cpu_power(struct sched_domain *sd, int cpu)
5215 unsigned long weight = sd->span_weight;
5216 unsigned long power = SCHED_POWER_SCALE;
5217 struct sched_group *sdg = sd->groups;
5219 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
5220 if (sched_feat(ARCH_POWER))
5221 power *= arch_scale_smt_power(sd, cpu);
5223 power *= default_scale_smt_power(sd, cpu);
5225 power >>= SCHED_POWER_SHIFT;
5228 sdg->sgp->power_orig = power;
5230 if (sched_feat(ARCH_POWER))
5231 power *= arch_scale_freq_power(sd, cpu);
5233 power *= default_scale_freq_power(sd, cpu);
5235 power >>= SCHED_POWER_SHIFT;
5237 power *= scale_rt_power(cpu);
5238 power >>= SCHED_POWER_SHIFT;
5243 cpu_rq(cpu)->cpu_power = power;
5244 sdg->sgp->power = power;
5247 void update_group_power(struct sched_domain *sd, int cpu)
5249 struct sched_domain *child = sd->child;
5250 struct sched_group *group, *sdg = sd->groups;
5251 unsigned long power, power_orig;
5252 unsigned long interval;
5254 interval = msecs_to_jiffies(sd->balance_interval);
5255 interval = clamp(interval, 1UL, max_load_balance_interval);
5256 sdg->sgp->next_update = jiffies + interval;
5259 update_cpu_power(sd, cpu);
5263 power_orig = power = 0;
5265 if (child->flags & SD_OVERLAP) {
5267 * SD_OVERLAP domains cannot assume that child groups
5268 * span the current group.
5271 for_each_cpu(cpu, sched_group_cpus(sdg)) {
5272 struct sched_group *sg = cpu_rq(cpu)->sd->groups;
5274 power_orig += sg->sgp->power_orig;
5275 power += sg->sgp->power;
5279 * !SD_OVERLAP domains can assume that child groups
5280 * span the current group.
5283 group = child->groups;
5285 power_orig += group->sgp->power_orig;
5286 power += group->sgp->power;
5287 group = group->next;
5288 } while (group != child->groups);
5291 sdg->sgp->power_orig = power_orig;
5292 sdg->sgp->power = power;
5296 * Try and fix up capacity for tiny siblings, this is needed when
5297 * things like SD_ASYM_PACKING need f_b_g to select another sibling
5298 * which on its own isn't powerful enough.
5300 * See update_sd_pick_busiest() and check_asym_packing().
5303 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
5306 * Only siblings can have significantly less than SCHED_POWER_SCALE
5308 if (!(sd->flags & SD_SHARE_CPUPOWER))
5312 * If ~90% of the cpu_power is still there, we're good.
5314 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
5321 * Group imbalance indicates (and tries to solve) the problem where balancing
5322 * groups is inadequate due to tsk_cpus_allowed() constraints.
5324 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
5325 * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
5328 * { 0 1 2 3 } { 4 5 6 7 }
5331 * If we were to balance group-wise we'd place two tasks in the first group and
5332 * two tasks in the second group. Clearly this is undesired as it will overload
5333 * cpu 3 and leave one of the cpus in the second group unused.
5335 * The current solution to this issue is detecting the skew in the first group
5336 * by noticing the lower domain failed to reach balance and had difficulty
5337 * moving tasks due to affinity constraints.
5339 * When this is so detected; this group becomes a candidate for busiest; see
5340 * update_sd_pick_busiest(). And calculcate_imbalance() and
5341 * find_busiest_group() avoid some of the usual balance conditions to allow it
5342 * to create an effective group imbalance.
5344 * This is a somewhat tricky proposition since the next run might not find the
5345 * group imbalance and decide the groups need to be balanced again. A most
5346 * subtle and fragile situation.
5349 static inline int sg_imbalanced(struct sched_group *group)
5351 return group->sgp->imbalance;
5355 * Compute the group capacity.
5357 * Avoid the issue where N*frac(smt_power) >= 1 creates 'phantom' cores by
5358 * first dividing out the smt factor and computing the actual number of cores
5359 * and limit power unit capacity with that.
5361 static inline int sg_capacity(struct lb_env *env, struct sched_group *group)
5363 unsigned int capacity, smt, cpus;
5364 unsigned int power, power_orig;
5366 power = group->sgp->power;
5367 power_orig = group->sgp->power_orig;
5368 cpus = group->group_weight;
5370 /* smt := ceil(cpus / power), assumes: 1 < smt_power < 2 */
5371 smt = DIV_ROUND_UP(SCHED_POWER_SCALE * cpus, power_orig);
5372 capacity = cpus / smt; /* cores */
5374 capacity = min_t(unsigned, capacity, DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE));
5376 capacity = fix_small_capacity(env->sd, group);
5382 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
5383 * @env: The load balancing environment.
5384 * @group: sched_group whose statistics are to be updated.
5385 * @load_idx: Load index of sched_domain of this_cpu for load calc.
5386 * @local_group: Does group contain this_cpu.
5387 * @sgs: variable to hold the statistics for this group.
5389 static inline void update_sg_lb_stats(struct lb_env *env,
5390 struct sched_group *group, int load_idx,
5391 int local_group, struct sg_lb_stats *sgs)
5393 unsigned long nr_running;
5397 memset(sgs, 0, sizeof(*sgs));
5399 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5400 struct rq *rq = cpu_rq(i);
5402 nr_running = rq->nr_running;
5404 /* Bias balancing toward cpus of our domain */
5406 load = target_load(i, load_idx);
5408 load = source_load(i, load_idx);
5410 sgs->group_load += load;
5411 sgs->sum_nr_running += nr_running;
5412 sgs->sum_weighted_load += weighted_cpuload(i);
5417 /* Adjust by relative CPU power of the group */
5418 sgs->group_power = group->sgp->power;
5419 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / sgs->group_power;
5421 if (sgs->sum_nr_running)
5422 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
5424 sgs->group_weight = group->group_weight;
5426 sgs->group_imb = sg_imbalanced(group);
5427 sgs->group_capacity = sg_capacity(env, group);
5429 if (sgs->group_capacity > sgs->sum_nr_running)
5430 sgs->group_has_capacity = 1;
5434 * update_sd_pick_busiest - return 1 on busiest group
5435 * @env: The load balancing environment.
5436 * @sds: sched_domain statistics
5437 * @sg: sched_group candidate to be checked for being the busiest
5438 * @sgs: sched_group statistics
5440 * Determine if @sg is a busier group than the previously selected
5443 * Return: %true if @sg is a busier group than the previously selected
5444 * busiest group. %false otherwise.
5446 static bool update_sd_pick_busiest(struct lb_env *env,
5447 struct sd_lb_stats *sds,
5448 struct sched_group *sg,
5449 struct sg_lb_stats *sgs)
5451 if (sgs->avg_load <= sds->busiest_stat.avg_load)
5454 if (sgs->sum_nr_running > sgs->group_capacity)
5461 * ASYM_PACKING needs to move all the work to the lowest
5462 * numbered CPUs in the group, therefore mark all groups
5463 * higher than ourself as busy.
5465 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
5466 env->dst_cpu < group_first_cpu(sg)) {
5470 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
5478 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
5479 * @env: The load balancing environment.
5480 * @balance: Should we balance.
5481 * @sds: variable to hold the statistics for this sched_domain.
5483 static inline void update_sd_lb_stats(struct lb_env *env,
5484 struct sd_lb_stats *sds)
5486 struct sched_domain *child = env->sd->child;
5487 struct sched_group *sg = env->sd->groups;
5488 struct sg_lb_stats tmp_sgs;
5489 int load_idx, prefer_sibling = 0;
5491 if (child && child->flags & SD_PREFER_SIBLING)
5494 load_idx = get_sd_load_idx(env->sd, env->idle);
5497 struct sg_lb_stats *sgs = &tmp_sgs;
5500 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
5503 sgs = &sds->local_stat;
5505 if (env->idle != CPU_NEWLY_IDLE ||
5506 time_after_eq(jiffies, sg->sgp->next_update))
5507 update_group_power(env->sd, env->dst_cpu);
5510 update_sg_lb_stats(env, sg, load_idx, local_group, sgs);
5516 * In case the child domain prefers tasks go to siblings
5517 * first, lower the sg capacity to one so that we'll try
5518 * and move all the excess tasks away. We lower the capacity
5519 * of a group only if the local group has the capacity to fit
5520 * these excess tasks, i.e. nr_running < group_capacity. The
5521 * extra check prevents the case where you always pull from the
5522 * heaviest group when it is already under-utilized (possible
5523 * with a large weight task outweighs the tasks on the system).
5525 if (prefer_sibling && sds->local &&
5526 sds->local_stat.group_has_capacity)
5527 sgs->group_capacity = min(sgs->group_capacity, 1U);
5529 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
5531 sds->busiest_stat = *sgs;
5535 /* Now, start updating sd_lb_stats */
5536 sds->total_load += sgs->group_load;
5537 sds->total_pwr += sgs->group_power;
5540 } while (sg != env->sd->groups);
5544 * check_asym_packing - Check to see if the group is packed into the
5547 * This is primarily intended to used at the sibling level. Some
5548 * cores like POWER7 prefer to use lower numbered SMT threads. In the
5549 * case of POWER7, it can move to lower SMT modes only when higher
5550 * threads are idle. When in lower SMT modes, the threads will
5551 * perform better since they share less core resources. Hence when we
5552 * have idle threads, we want them to be the higher ones.
5554 * This packing function is run on idle threads. It checks to see if
5555 * the busiest CPU in this domain (core in the P7 case) has a higher
5556 * CPU number than the packing function is being run on. Here we are
5557 * assuming lower CPU number will be equivalent to lower a SMT thread
5560 * Return: 1 when packing is required and a task should be moved to
5561 * this CPU. The amount of the imbalance is returned in *imbalance.
5563 * @env: The load balancing environment.
5564 * @sds: Statistics of the sched_domain which is to be packed
5566 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
5570 if (!(env->sd->flags & SD_ASYM_PACKING))
5576 busiest_cpu = group_first_cpu(sds->busiest);
5577 if (env->dst_cpu > busiest_cpu)
5580 env->imbalance = DIV_ROUND_CLOSEST(
5581 sds->busiest_stat.avg_load * sds->busiest_stat.group_power,
5588 * fix_small_imbalance - Calculate the minor imbalance that exists
5589 * amongst the groups of a sched_domain, during
5591 * @env: The load balancing environment.
5592 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
5595 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5597 unsigned long tmp, pwr_now = 0, pwr_move = 0;
5598 unsigned int imbn = 2;
5599 unsigned long scaled_busy_load_per_task;
5600 struct sg_lb_stats *local, *busiest;
5602 local = &sds->local_stat;
5603 busiest = &sds->busiest_stat;
5605 if (!local->sum_nr_running)
5606 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
5607 else if (busiest->load_per_task > local->load_per_task)
5610 scaled_busy_load_per_task =
5611 (busiest->load_per_task * SCHED_POWER_SCALE) /
5612 busiest->group_power;
5614 if (busiest->avg_load + scaled_busy_load_per_task >=
5615 local->avg_load + (scaled_busy_load_per_task * imbn)) {
5616 env->imbalance = busiest->load_per_task;
5621 * OK, we don't have enough imbalance to justify moving tasks,
5622 * however we may be able to increase total CPU power used by
5626 pwr_now += busiest->group_power *
5627 min(busiest->load_per_task, busiest->avg_load);
5628 pwr_now += local->group_power *
5629 min(local->load_per_task, local->avg_load);
5630 pwr_now /= SCHED_POWER_SCALE;
5632 /* Amount of load we'd subtract */
5633 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5634 busiest->group_power;
5635 if (busiest->avg_load > tmp) {
5636 pwr_move += busiest->group_power *
5637 min(busiest->load_per_task,
5638 busiest->avg_load - tmp);
5641 /* Amount of load we'd add */
5642 if (busiest->avg_load * busiest->group_power <
5643 busiest->load_per_task * SCHED_POWER_SCALE) {
5644 tmp = (busiest->avg_load * busiest->group_power) /
5647 tmp = (busiest->load_per_task * SCHED_POWER_SCALE) /
5650 pwr_move += local->group_power *
5651 min(local->load_per_task, local->avg_load + tmp);
5652 pwr_move /= SCHED_POWER_SCALE;
5654 /* Move if we gain throughput */
5655 if (pwr_move > pwr_now)
5656 env->imbalance = busiest->load_per_task;
5660 * calculate_imbalance - Calculate the amount of imbalance present within the
5661 * groups of a given sched_domain during load balance.
5662 * @env: load balance environment
5663 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
5665 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
5667 unsigned long max_pull, load_above_capacity = ~0UL;
5668 struct sg_lb_stats *local, *busiest;
5670 local = &sds->local_stat;
5671 busiest = &sds->busiest_stat;
5673 if (busiest->group_imb) {
5675 * In the group_imb case we cannot rely on group-wide averages
5676 * to ensure cpu-load equilibrium, look at wider averages. XXX
5678 busiest->load_per_task =
5679 min(busiest->load_per_task, sds->avg_load);
5683 * In the presence of smp nice balancing, certain scenarios can have
5684 * max load less than avg load(as we skip the groups at or below
5685 * its cpu_power, while calculating max_load..)
5687 if (busiest->avg_load <= sds->avg_load ||
5688 local->avg_load >= sds->avg_load) {
5690 return fix_small_imbalance(env, sds);
5693 if (!busiest->group_imb) {
5695 * Don't want to pull so many tasks that a group would go idle.
5696 * Except of course for the group_imb case, since then we might
5697 * have to drop below capacity to reach cpu-load equilibrium.
5699 load_above_capacity =
5700 (busiest->sum_nr_running - busiest->group_capacity);
5702 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
5703 load_above_capacity /= busiest->group_power;
5707 * We're trying to get all the cpus to the average_load, so we don't
5708 * want to push ourselves above the average load, nor do we wish to
5709 * reduce the max loaded cpu below the average load. At the same time,
5710 * we also don't want to reduce the group load below the group capacity
5711 * (so that we can implement power-savings policies etc). Thus we look
5712 * for the minimum possible imbalance.
5714 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
5716 /* How much load to actually move to equalise the imbalance */
5717 env->imbalance = min(
5718 max_pull * busiest->group_power,
5719 (sds->avg_load - local->avg_load) * local->group_power
5720 ) / SCHED_POWER_SCALE;
5723 * if *imbalance is less than the average load per runnable task
5724 * there is no guarantee that any tasks will be moved so we'll have
5725 * a think about bumping its value to force at least one task to be
5728 if (env->imbalance < busiest->load_per_task)
5729 return fix_small_imbalance(env, sds);
5732 /******* find_busiest_group() helpers end here *********************/
5735 * find_busiest_group - Returns the busiest group within the sched_domain
5736 * if there is an imbalance. If there isn't an imbalance, and
5737 * the user has opted for power-savings, it returns a group whose
5738 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
5739 * such a group exists.
5741 * Also calculates the amount of weighted load which should be moved
5742 * to restore balance.
5744 * @env: The load balancing environment.
5746 * Return: - The busiest group if imbalance exists.
5747 * - If no imbalance and user has opted for power-savings balance,
5748 * return the least loaded group whose CPUs can be
5749 * put to idle by rebalancing its tasks onto our group.
5751 static struct sched_group *find_busiest_group(struct lb_env *env)
5753 struct sg_lb_stats *local, *busiest;
5754 struct sd_lb_stats sds;
5756 init_sd_lb_stats(&sds);
5759 * Compute the various statistics relavent for load balancing at
5762 update_sd_lb_stats(env, &sds);
5763 local = &sds.local_stat;
5764 busiest = &sds.busiest_stat;
5766 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
5767 check_asym_packing(env, &sds))
5770 /* There is no busy sibling group to pull tasks from */
5771 if (!sds.busiest || busiest->sum_nr_running == 0)
5774 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
5777 * If the busiest group is imbalanced the below checks don't
5778 * work because they assume all things are equal, which typically
5779 * isn't true due to cpus_allowed constraints and the like.
5781 if (busiest->group_imb)
5784 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
5785 if (env->idle == CPU_NEWLY_IDLE && local->group_has_capacity &&
5786 !busiest->group_has_capacity)
5790 * If the local group is more busy than the selected busiest group
5791 * don't try and pull any tasks.
5793 if (local->avg_load >= busiest->avg_load)
5797 * Don't pull any tasks if this group is already above the domain
5800 if (local->avg_load >= sds.avg_load)
5803 if (env->idle == CPU_IDLE) {
5805 * This cpu is idle. If the busiest group load doesn't
5806 * have more tasks than the number of available cpu's and
5807 * there is no imbalance between this and busiest group
5808 * wrt to idle cpu's, it is balanced.
5810 if ((local->idle_cpus < busiest->idle_cpus) &&
5811 busiest->sum_nr_running <= busiest->group_weight)
5815 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
5816 * imbalance_pct to be conservative.
5818 if (100 * busiest->avg_load <=
5819 env->sd->imbalance_pct * local->avg_load)
5824 /* Looks like there is an imbalance. Compute it */
5825 calculate_imbalance(env, &sds);
5834 * find_busiest_queue - find the busiest runqueue among the cpus in group.
5836 static struct rq *find_busiest_queue(struct lb_env *env,
5837 struct sched_group *group)
5839 struct rq *busiest = NULL, *rq;
5840 unsigned long busiest_load = 0, busiest_power = 1;
5843 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
5844 unsigned long power = power_of(i);
5845 unsigned long capacity = DIV_ROUND_CLOSEST(power,
5850 capacity = fix_small_capacity(env->sd, group);
5853 wl = weighted_cpuload(i);
5856 * When comparing with imbalance, use weighted_cpuload()
5857 * which is not scaled with the cpu power.
5859 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
5863 * For the load comparisons with the other cpu's, consider
5864 * the weighted_cpuload() scaled with the cpu power, so that
5865 * the load can be moved away from the cpu that is potentially
5866 * running at a lower capacity.
5868 * Thus we're looking for max(wl_i / power_i), crosswise
5869 * multiplication to rid ourselves of the division works out
5870 * to: wl_i * power_j > wl_j * power_i; where j is our
5873 if (wl * busiest_power > busiest_load * power) {
5875 busiest_power = power;
5884 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
5885 * so long as it is large enough.
5887 #define MAX_PINNED_INTERVAL 512
5889 /* Working cpumask for load_balance and load_balance_newidle. */
5890 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5892 static int need_active_balance(struct lb_env *env)
5894 struct sched_domain *sd = env->sd;
5896 if (env->idle == CPU_NEWLY_IDLE) {
5899 * ASYM_PACKING needs to force migrate tasks from busy but
5900 * higher numbered CPUs in order to pack all tasks in the
5901 * lowest numbered CPUs.
5903 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
5907 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
5910 static int active_load_balance_cpu_stop(void *data);
5912 static int should_we_balance(struct lb_env *env)
5914 struct sched_group *sg = env->sd->groups;
5915 struct cpumask *sg_cpus, *sg_mask;
5916 int cpu, balance_cpu = -1;
5919 * In the newly idle case, we will allow all the cpu's
5920 * to do the newly idle load balance.
5922 if (env->idle == CPU_NEWLY_IDLE)
5925 sg_cpus = sched_group_cpus(sg);
5926 sg_mask = sched_group_mask(sg);
5927 /* Try to find first idle cpu */
5928 for_each_cpu_and(cpu, sg_cpus, env->cpus) {
5929 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
5936 if (balance_cpu == -1)
5937 balance_cpu = group_balance_cpu(sg);
5940 * First idle cpu or the first cpu(busiest) in this sched group
5941 * is eligible for doing load balancing at this and above domains.
5943 return balance_cpu == env->dst_cpu;
5947 * Check this_cpu to ensure it is balanced within domain. Attempt to move
5948 * tasks if there is an imbalance.
5950 static int load_balance(int this_cpu, struct rq *this_rq,
5951 struct sched_domain *sd, enum cpu_idle_type idle,
5952 int *continue_balancing)
5954 int ld_moved, cur_ld_moved, active_balance = 0;
5955 struct sched_domain *sd_parent = sd->parent;
5956 struct sched_group *group;
5958 unsigned long flags;
5959 struct cpumask *cpus = __get_cpu_var(load_balance_mask);
5961 struct lb_env env = {
5963 .dst_cpu = this_cpu,
5965 .dst_grpmask = sched_group_cpus(sd->groups),
5967 .loop_break = sched_nr_migrate_break,
5972 * For NEWLY_IDLE load_balancing, we don't need to consider
5973 * other cpus in our group
5975 if (idle == CPU_NEWLY_IDLE)
5976 env.dst_grpmask = NULL;
5978 cpumask_copy(cpus, cpu_active_mask);
5980 schedstat_inc(sd, lb_count[idle]);
5983 if (!should_we_balance(&env)) {
5984 *continue_balancing = 0;
5988 group = find_busiest_group(&env);
5990 schedstat_inc(sd, lb_nobusyg[idle]);
5994 busiest = find_busiest_queue(&env, group);
5996 schedstat_inc(sd, lb_nobusyq[idle]);
6000 BUG_ON(busiest == env.dst_rq);
6002 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
6005 if (busiest->nr_running > 1) {
6007 * Attempt to move tasks. If find_busiest_group has found
6008 * an imbalance but busiest->nr_running <= 1, the group is
6009 * still unbalanced. ld_moved simply stays zero, so it is
6010 * correctly treated as an imbalance.
6012 env.flags |= LBF_ALL_PINNED;
6013 env.src_cpu = busiest->cpu;
6014 env.src_rq = busiest;
6015 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
6018 local_irq_save(flags);
6019 double_rq_lock(env.dst_rq, busiest);
6022 * cur_ld_moved - load moved in current iteration
6023 * ld_moved - cumulative load moved across iterations
6025 cur_ld_moved = move_tasks(&env);
6026 ld_moved += cur_ld_moved;
6027 double_rq_unlock(env.dst_rq, busiest);
6028 local_irq_restore(flags);
6031 * some other cpu did the load balance for us.
6033 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
6034 resched_cpu(env.dst_cpu);
6036 if (env.flags & LBF_NEED_BREAK) {
6037 env.flags &= ~LBF_NEED_BREAK;
6042 * Revisit (affine) tasks on src_cpu that couldn't be moved to
6043 * us and move them to an alternate dst_cpu in our sched_group
6044 * where they can run. The upper limit on how many times we
6045 * iterate on same src_cpu is dependent on number of cpus in our
6048 * This changes load balance semantics a bit on who can move
6049 * load to a given_cpu. In addition to the given_cpu itself
6050 * (or a ilb_cpu acting on its behalf where given_cpu is
6051 * nohz-idle), we now have balance_cpu in a position to move
6052 * load to given_cpu. In rare situations, this may cause
6053 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
6054 * _independently_ and at _same_ time to move some load to
6055 * given_cpu) causing exceess load to be moved to given_cpu.
6056 * This however should not happen so much in practice and
6057 * moreover subsequent load balance cycles should correct the
6058 * excess load moved.
6060 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
6062 /* Prevent to re-select dst_cpu via env's cpus */
6063 cpumask_clear_cpu(env.dst_cpu, env.cpus);
6065 env.dst_rq = cpu_rq(env.new_dst_cpu);
6066 env.dst_cpu = env.new_dst_cpu;
6067 env.flags &= ~LBF_DST_PINNED;
6069 env.loop_break = sched_nr_migrate_break;
6072 * Go back to "more_balance" rather than "redo" since we
6073 * need to continue with same src_cpu.
6079 * We failed to reach balance because of affinity.
6082 int *group_imbalance = &sd_parent->groups->sgp->imbalance;
6084 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
6085 *group_imbalance = 1;
6086 } else if (*group_imbalance)
6087 *group_imbalance = 0;
6090 /* All tasks on this runqueue were pinned by CPU affinity */
6091 if (unlikely(env.flags & LBF_ALL_PINNED)) {
6092 cpumask_clear_cpu(cpu_of(busiest), cpus);
6093 if (!cpumask_empty(cpus)) {
6095 env.loop_break = sched_nr_migrate_break;
6103 schedstat_inc(sd, lb_failed[idle]);
6105 * Increment the failure counter only on periodic balance.
6106 * We do not want newidle balance, which can be very
6107 * frequent, pollute the failure counter causing
6108 * excessive cache_hot migrations and active balances.
6110 if (idle != CPU_NEWLY_IDLE)
6111 sd->nr_balance_failed++;
6113 if (need_active_balance(&env)) {
6114 raw_spin_lock_irqsave(&busiest->lock, flags);
6116 /* don't kick the active_load_balance_cpu_stop,
6117 * if the curr task on busiest cpu can't be
6120 if (!cpumask_test_cpu(this_cpu,
6121 tsk_cpus_allowed(busiest->curr))) {
6122 raw_spin_unlock_irqrestore(&busiest->lock,
6124 env.flags |= LBF_ALL_PINNED;
6125 goto out_one_pinned;
6129 * ->active_balance synchronizes accesses to
6130 * ->active_balance_work. Once set, it's cleared
6131 * only after active load balance is finished.
6133 if (!busiest->active_balance) {
6134 busiest->active_balance = 1;
6135 busiest->push_cpu = this_cpu;
6138 raw_spin_unlock_irqrestore(&busiest->lock, flags);
6140 if (active_balance) {
6141 stop_one_cpu_nowait(cpu_of(busiest),
6142 active_load_balance_cpu_stop, busiest,
6143 &busiest->active_balance_work);
6147 * We've kicked active balancing, reset the failure
6150 sd->nr_balance_failed = sd->cache_nice_tries+1;
6153 sd->nr_balance_failed = 0;
6155 if (likely(!active_balance)) {
6156 /* We were unbalanced, so reset the balancing interval */
6157 sd->balance_interval = sd->min_interval;
6160 * If we've begun active balancing, start to back off. This
6161 * case may not be covered by the all_pinned logic if there
6162 * is only 1 task on the busy runqueue (because we don't call
6165 if (sd->balance_interval < sd->max_interval)
6166 sd->balance_interval *= 2;
6172 schedstat_inc(sd, lb_balanced[idle]);
6174 sd->nr_balance_failed = 0;
6177 /* tune up the balancing interval */
6178 if (((env.flags & LBF_ALL_PINNED) &&
6179 sd->balance_interval < MAX_PINNED_INTERVAL) ||
6180 (sd->balance_interval < sd->max_interval))
6181 sd->balance_interval *= 2;
6189 * idle_balance is called by schedule() if this_cpu is about to become
6190 * idle. Attempts to pull tasks from other CPUs.
6192 void idle_balance(int this_cpu, struct rq *this_rq)
6194 struct sched_domain *sd;
6195 int pulled_task = 0;
6196 unsigned long next_balance = jiffies + HZ;
6199 this_rq->idle_stamp = rq_clock(this_rq);
6201 if (this_rq->avg_idle < sysctl_sched_migration_cost)
6205 * Drop the rq->lock, but keep IRQ/preempt disabled.
6207 raw_spin_unlock(&this_rq->lock);
6209 update_blocked_averages(this_cpu);
6211 for_each_domain(this_cpu, sd) {
6212 unsigned long interval;
6213 int continue_balancing = 1;
6214 u64 t0, domain_cost;
6216 if (!(sd->flags & SD_LOAD_BALANCE))
6219 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
6222 if (sd->flags & SD_BALANCE_NEWIDLE) {
6223 t0 = sched_clock_cpu(this_cpu);
6225 /* If we've pulled tasks over stop searching: */
6226 pulled_task = load_balance(this_cpu, this_rq,
6228 &continue_balancing);
6230 domain_cost = sched_clock_cpu(this_cpu) - t0;
6231 if (domain_cost > sd->max_newidle_lb_cost)
6232 sd->max_newidle_lb_cost = domain_cost;
6234 curr_cost += domain_cost;
6237 interval = msecs_to_jiffies(sd->balance_interval);
6238 if (time_after(next_balance, sd->last_balance + interval))
6239 next_balance = sd->last_balance + interval;
6241 this_rq->idle_stamp = 0;
6247 raw_spin_lock(&this_rq->lock);
6249 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
6251 * We are going idle. next_balance may be set based on
6252 * a busy processor. So reset next_balance.
6254 this_rq->next_balance = next_balance;
6257 if (curr_cost > this_rq->max_idle_balance_cost)
6258 this_rq->max_idle_balance_cost = curr_cost;
6262 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
6263 * running tasks off the busiest CPU onto idle CPUs. It requires at
6264 * least 1 task to be running on each physical CPU where possible, and
6265 * avoids physical / logical imbalances.
6267 static int active_load_balance_cpu_stop(void *data)
6269 struct rq *busiest_rq = data;
6270 int busiest_cpu = cpu_of(busiest_rq);
6271 int target_cpu = busiest_rq->push_cpu;
6272 struct rq *target_rq = cpu_rq(target_cpu);
6273 struct sched_domain *sd;
6275 raw_spin_lock_irq(&busiest_rq->lock);
6277 /* make sure the requested cpu hasn't gone down in the meantime */
6278 if (unlikely(busiest_cpu != smp_processor_id() ||
6279 !busiest_rq->active_balance))
6282 /* Is there any task to move? */
6283 if (busiest_rq->nr_running <= 1)
6287 * This condition is "impossible", if it occurs
6288 * we need to fix it. Originally reported by
6289 * Bjorn Helgaas on a 128-cpu setup.
6291 BUG_ON(busiest_rq == target_rq);
6293 /* move a task from busiest_rq to target_rq */
6294 double_lock_balance(busiest_rq, target_rq);
6296 /* Search for an sd spanning us and the target CPU. */
6298 for_each_domain(target_cpu, sd) {
6299 if ((sd->flags & SD_LOAD_BALANCE) &&
6300 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
6305 struct lb_env env = {
6307 .dst_cpu = target_cpu,
6308 .dst_rq = target_rq,
6309 .src_cpu = busiest_rq->cpu,
6310 .src_rq = busiest_rq,
6314 schedstat_inc(sd, alb_count);
6316 if (move_one_task(&env))
6317 schedstat_inc(sd, alb_pushed);
6319 schedstat_inc(sd, alb_failed);
6322 double_unlock_balance(busiest_rq, target_rq);
6324 busiest_rq->active_balance = 0;
6325 raw_spin_unlock_irq(&busiest_rq->lock);
6329 #ifdef CONFIG_NO_HZ_COMMON
6331 * idle load balancing details
6332 * - When one of the busy CPUs notice that there may be an idle rebalancing
6333 * needed, they will kick the idle load balancer, which then does idle
6334 * load balancing for all the idle CPUs.
6337 cpumask_var_t idle_cpus_mask;
6339 unsigned long next_balance; /* in jiffy units */
6340 } nohz ____cacheline_aligned;
6342 static inline int find_new_ilb(int call_cpu)
6344 int ilb = cpumask_first(nohz.idle_cpus_mask);
6346 if (ilb < nr_cpu_ids && idle_cpu(ilb))
6353 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
6354 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
6355 * CPU (if there is one).
6357 static void nohz_balancer_kick(int cpu)
6361 nohz.next_balance++;
6363 ilb_cpu = find_new_ilb(cpu);
6365 if (ilb_cpu >= nr_cpu_ids)
6368 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
6371 * Use smp_send_reschedule() instead of resched_cpu().
6372 * This way we generate a sched IPI on the target cpu which
6373 * is idle. And the softirq performing nohz idle load balance
6374 * will be run before returning from the IPI.
6376 smp_send_reschedule(ilb_cpu);
6380 static inline void nohz_balance_exit_idle(int cpu)
6382 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
6383 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
6384 atomic_dec(&nohz.nr_cpus);
6385 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6389 static inline void set_cpu_sd_state_busy(void)
6391 struct sched_domain *sd;
6394 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6396 if (!sd || !sd->nohz_idle)
6400 for (; sd; sd = sd->parent)
6401 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
6406 void set_cpu_sd_state_idle(void)
6408 struct sched_domain *sd;
6411 sd = rcu_dereference_check_sched_domain(this_rq()->sd);
6413 if (!sd || sd->nohz_idle)
6417 for (; sd; sd = sd->parent)
6418 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
6424 * This routine will record that the cpu is going idle with tick stopped.
6425 * This info will be used in performing idle load balancing in the future.
6427 void nohz_balance_enter_idle(int cpu)
6430 * If this cpu is going down, then nothing needs to be done.
6432 if (!cpu_active(cpu))
6435 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
6438 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
6439 atomic_inc(&nohz.nr_cpus);
6440 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
6443 static int sched_ilb_notifier(struct notifier_block *nfb,
6444 unsigned long action, void *hcpu)
6446 switch (action & ~CPU_TASKS_FROZEN) {
6448 nohz_balance_exit_idle(smp_processor_id());
6456 static DEFINE_SPINLOCK(balancing);
6459 * Scale the max load_balance interval with the number of CPUs in the system.
6460 * This trades load-balance latency on larger machines for less cross talk.
6462 void update_max_interval(void)
6464 max_load_balance_interval = HZ*num_online_cpus()/10;
6468 * It checks each scheduling domain to see if it is due to be balanced,
6469 * and initiates a balancing operation if so.
6471 * Balancing parameters are set up in init_sched_domains.
6473 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
6475 int continue_balancing = 1;
6476 struct rq *rq = cpu_rq(cpu);
6477 unsigned long interval;
6478 struct sched_domain *sd;
6479 /* Earliest time when we have to do rebalance again */
6480 unsigned long next_balance = jiffies + 60*HZ;
6481 int update_next_balance = 0;
6482 int need_serialize, need_decay = 0;
6485 update_blocked_averages(cpu);
6488 for_each_domain(cpu, sd) {
6490 * Decay the newidle max times here because this is a regular
6491 * visit to all the domains. Decay ~1% per second.
6493 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
6494 sd->max_newidle_lb_cost =
6495 (sd->max_newidle_lb_cost * 253) / 256;
6496 sd->next_decay_max_lb_cost = jiffies + HZ;
6499 max_cost += sd->max_newidle_lb_cost;
6501 if (!(sd->flags & SD_LOAD_BALANCE))
6505 * Stop the load balance at this level. There is another
6506 * CPU in our sched group which is doing load balancing more
6509 if (!continue_balancing) {
6515 interval = sd->balance_interval;
6516 if (idle != CPU_IDLE)
6517 interval *= sd->busy_factor;
6519 /* scale ms to jiffies */
6520 interval = msecs_to_jiffies(interval);
6521 interval = clamp(interval, 1UL, max_load_balance_interval);
6523 need_serialize = sd->flags & SD_SERIALIZE;
6525 if (need_serialize) {
6526 if (!spin_trylock(&balancing))
6530 if (time_after_eq(jiffies, sd->last_balance + interval)) {
6531 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
6533 * The LBF_DST_PINNED logic could have changed
6534 * env->dst_cpu, so we can't know our idle
6535 * state even if we migrated tasks. Update it.
6537 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
6539 sd->last_balance = jiffies;
6542 spin_unlock(&balancing);
6544 if (time_after(next_balance, sd->last_balance + interval)) {
6545 next_balance = sd->last_balance + interval;
6546 update_next_balance = 1;
6551 * Ensure the rq-wide value also decays but keep it at a
6552 * reasonable floor to avoid funnies with rq->avg_idle.
6554 rq->max_idle_balance_cost =
6555 max((u64)sysctl_sched_migration_cost, max_cost);
6560 * next_balance will be updated only when there is a need.
6561 * When the cpu is attached to null domain for ex, it will not be
6564 if (likely(update_next_balance))
6565 rq->next_balance = next_balance;
6568 #ifdef CONFIG_NO_HZ_COMMON
6570 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
6571 * rebalancing for all the cpus for whom scheduler ticks are stopped.
6573 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
6575 struct rq *this_rq = cpu_rq(this_cpu);
6579 if (idle != CPU_IDLE ||
6580 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
6583 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
6584 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
6588 * If this cpu gets work to do, stop the load balancing
6589 * work being done for other cpus. Next load
6590 * balancing owner will pick it up.
6595 rq = cpu_rq(balance_cpu);
6597 raw_spin_lock_irq(&rq->lock);
6598 update_rq_clock(rq);
6599 update_idle_cpu_load(rq);
6600 raw_spin_unlock_irq(&rq->lock);
6602 rebalance_domains(balance_cpu, CPU_IDLE);
6604 if (time_after(this_rq->next_balance, rq->next_balance))
6605 this_rq->next_balance = rq->next_balance;
6607 nohz.next_balance = this_rq->next_balance;
6609 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
6613 * Current heuristic for kicking the idle load balancer in the presence
6614 * of an idle cpu is the system.
6615 * - This rq has more than one task.
6616 * - At any scheduler domain level, this cpu's scheduler group has multiple
6617 * busy cpu's exceeding the group's power.
6618 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
6619 * domain span are idle.
6621 static inline int nohz_kick_needed(struct rq *rq, int cpu)
6623 unsigned long now = jiffies;
6624 struct sched_domain *sd;
6626 if (unlikely(idle_cpu(cpu)))
6630 * We may be recently in ticked or tickless idle mode. At the first
6631 * busy tick after returning from idle, we will update the busy stats.
6633 set_cpu_sd_state_busy();
6634 nohz_balance_exit_idle(cpu);
6637 * None are in tickless mode and hence no need for NOHZ idle load
6640 if (likely(!atomic_read(&nohz.nr_cpus)))
6643 if (time_before(now, nohz.next_balance))
6646 if (rq->nr_running >= 2)
6650 for_each_domain(cpu, sd) {
6651 struct sched_group *sg = sd->groups;
6652 struct sched_group_power *sgp = sg->sgp;
6653 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
6655 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
6656 goto need_kick_unlock;
6658 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
6659 && (cpumask_first_and(nohz.idle_cpus_mask,
6660 sched_domain_span(sd)) < cpu))
6661 goto need_kick_unlock;
6663 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
6675 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
6679 * run_rebalance_domains is triggered when needed from the scheduler tick.
6680 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
6682 static void run_rebalance_domains(struct softirq_action *h)
6684 int this_cpu = smp_processor_id();
6685 struct rq *this_rq = cpu_rq(this_cpu);
6686 enum cpu_idle_type idle = this_rq->idle_balance ?
6687 CPU_IDLE : CPU_NOT_IDLE;
6689 rebalance_domains(this_cpu, idle);
6692 * If this cpu has a pending nohz_balance_kick, then do the
6693 * balancing on behalf of the other idle cpus whose ticks are
6696 nohz_idle_balance(this_cpu, idle);
6699 static inline int on_null_domain(int cpu)
6701 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
6705 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
6707 void trigger_load_balance(struct rq *rq, int cpu)
6709 /* Don't need to rebalance while attached to NULL domain */
6710 if (time_after_eq(jiffies, rq->next_balance) &&
6711 likely(!on_null_domain(cpu)))
6712 raise_softirq(SCHED_SOFTIRQ);
6713 #ifdef CONFIG_NO_HZ_COMMON
6714 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
6715 nohz_balancer_kick(cpu);
6719 static void rq_online_fair(struct rq *rq)
6724 static void rq_offline_fair(struct rq *rq)
6728 /* Ensure any throttled groups are reachable by pick_next_task */
6729 unthrottle_offline_cfs_rqs(rq);
6732 #endif /* CONFIG_SMP */
6735 * scheduler tick hitting a task of our scheduling class:
6737 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
6739 struct cfs_rq *cfs_rq;
6740 struct sched_entity *se = &curr->se;
6742 for_each_sched_entity(se) {
6743 cfs_rq = cfs_rq_of(se);
6744 entity_tick(cfs_rq, se, queued);
6747 if (numabalancing_enabled)
6748 task_tick_numa(rq, curr);
6750 update_rq_runnable_avg(rq, 1);
6754 * called on fork with the child task as argument from the parent's context
6755 * - child not yet on the tasklist
6756 * - preemption disabled
6758 static void task_fork_fair(struct task_struct *p)
6760 struct cfs_rq *cfs_rq;
6761 struct sched_entity *se = &p->se, *curr;
6762 int this_cpu = smp_processor_id();
6763 struct rq *rq = this_rq();
6764 unsigned long flags;
6766 raw_spin_lock_irqsave(&rq->lock, flags);
6768 update_rq_clock(rq);
6770 cfs_rq = task_cfs_rq(current);
6771 curr = cfs_rq->curr;
6774 * Not only the cpu but also the task_group of the parent might have
6775 * been changed after parent->se.parent,cfs_rq were copied to
6776 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
6777 * of child point to valid ones.
6780 __set_task_cpu(p, this_cpu);
6783 update_curr(cfs_rq);
6786 se->vruntime = curr->vruntime;
6787 place_entity(cfs_rq, se, 1);
6789 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
6791 * Upon rescheduling, sched_class::put_prev_task() will place
6792 * 'current' within the tree based on its new key value.
6794 swap(curr->vruntime, se->vruntime);
6795 resched_task(rq->curr);
6798 se->vruntime -= cfs_rq->min_vruntime;
6800 raw_spin_unlock_irqrestore(&rq->lock, flags);
6804 * Priority of the task has changed. Check to see if we preempt
6808 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
6814 * Reschedule if we are currently running on this runqueue and
6815 * our priority decreased, or if we are not currently running on
6816 * this runqueue and our priority is higher than the current's
6818 if (rq->curr == p) {
6819 if (p->prio > oldprio)
6820 resched_task(rq->curr);
6822 check_preempt_curr(rq, p, 0);
6825 static void switched_from_fair(struct rq *rq, struct task_struct *p)
6827 struct sched_entity *se = &p->se;
6828 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6831 * Ensure the task's vruntime is normalized, so that when its
6832 * switched back to the fair class the enqueue_entity(.flags=0) will
6833 * do the right thing.
6835 * If it was on_rq, then the dequeue_entity(.flags=0) will already
6836 * have normalized the vruntime, if it was !on_rq, then only when
6837 * the task is sleeping will it still have non-normalized vruntime.
6839 if (!se->on_rq && p->state != TASK_RUNNING) {
6841 * Fix up our vruntime so that the current sleep doesn't
6842 * cause 'unlimited' sleep bonus.
6844 place_entity(cfs_rq, se, 0);
6845 se->vruntime -= cfs_rq->min_vruntime;
6850 * Remove our load from contribution when we leave sched_fair
6851 * and ensure we don't carry in an old decay_count if we
6854 if (se->avg.decay_count) {
6855 __synchronize_entity_decay(se);
6856 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
6862 * We switched to the sched_fair class.
6864 static void switched_to_fair(struct rq *rq, struct task_struct *p)
6870 * We were most likely switched from sched_rt, so
6871 * kick off the schedule if running, otherwise just see
6872 * if we can still preempt the current task.
6875 resched_task(rq->curr);
6877 check_preempt_curr(rq, p, 0);
6880 /* Account for a task changing its policy or group.
6882 * This routine is mostly called to set cfs_rq->curr field when a task
6883 * migrates between groups/classes.
6885 static void set_curr_task_fair(struct rq *rq)
6887 struct sched_entity *se = &rq->curr->se;
6889 for_each_sched_entity(se) {
6890 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6892 set_next_entity(cfs_rq, se);
6893 /* ensure bandwidth has been allocated on our new cfs_rq */
6894 account_cfs_rq_runtime(cfs_rq, 0);
6898 void init_cfs_rq(struct cfs_rq *cfs_rq)
6900 cfs_rq->tasks_timeline = RB_ROOT;
6901 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
6902 #ifndef CONFIG_64BIT
6903 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
6906 atomic64_set(&cfs_rq->decay_counter, 1);
6907 atomic_long_set(&cfs_rq->removed_load, 0);
6911 #ifdef CONFIG_FAIR_GROUP_SCHED
6912 static void task_move_group_fair(struct task_struct *p, int on_rq)
6914 struct cfs_rq *cfs_rq;
6916 * If the task was not on the rq at the time of this cgroup movement
6917 * it must have been asleep, sleeping tasks keep their ->vruntime
6918 * absolute on their old rq until wakeup (needed for the fair sleeper
6919 * bonus in place_entity()).
6921 * If it was on the rq, we've just 'preempted' it, which does convert
6922 * ->vruntime to a relative base.
6924 * Make sure both cases convert their relative position when migrating
6925 * to another cgroup's rq. This does somewhat interfere with the
6926 * fair sleeper stuff for the first placement, but who cares.
6929 * When !on_rq, vruntime of the task has usually NOT been normalized.
6930 * But there are some cases where it has already been normalized:
6932 * - Moving a forked child which is waiting for being woken up by
6933 * wake_up_new_task().
6934 * - Moving a task which has been woken up by try_to_wake_up() and
6935 * waiting for actually being woken up by sched_ttwu_pending().
6937 * To prevent boost or penalty in the new cfs_rq caused by delta
6938 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
6940 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
6944 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
6945 set_task_rq(p, task_cpu(p));
6947 cfs_rq = cfs_rq_of(&p->se);
6948 p->se.vruntime += cfs_rq->min_vruntime;
6951 * migrate_task_rq_fair() will have removed our previous
6952 * contribution, but we must synchronize for ongoing future
6955 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
6956 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
6961 void free_fair_sched_group(struct task_group *tg)
6965 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
6967 for_each_possible_cpu(i) {
6969 kfree(tg->cfs_rq[i]);
6978 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
6980 struct cfs_rq *cfs_rq;
6981 struct sched_entity *se;
6984 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
6987 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
6991 tg->shares = NICE_0_LOAD;
6993 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
6995 for_each_possible_cpu(i) {
6996 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
6997 GFP_KERNEL, cpu_to_node(i));
7001 se = kzalloc_node(sizeof(struct sched_entity),
7002 GFP_KERNEL, cpu_to_node(i));
7006 init_cfs_rq(cfs_rq);
7007 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
7018 void unregister_fair_sched_group(struct task_group *tg, int cpu)
7020 struct rq *rq = cpu_rq(cpu);
7021 unsigned long flags;
7024 * Only empty task groups can be destroyed; so we can speculatively
7025 * check on_list without danger of it being re-added.
7027 if (!tg->cfs_rq[cpu]->on_list)
7030 raw_spin_lock_irqsave(&rq->lock, flags);
7031 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
7032 raw_spin_unlock_irqrestore(&rq->lock, flags);
7035 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7036 struct sched_entity *se, int cpu,
7037 struct sched_entity *parent)
7039 struct rq *rq = cpu_rq(cpu);
7043 init_cfs_rq_runtime(cfs_rq);
7045 tg->cfs_rq[cpu] = cfs_rq;
7048 /* se could be NULL for root_task_group */
7053 se->cfs_rq = &rq->cfs;
7055 se->cfs_rq = parent->my_q;
7058 update_load_set(&se->load, 0);
7059 se->parent = parent;
7062 static DEFINE_MUTEX(shares_mutex);
7064 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
7067 unsigned long flags;
7070 * We can't change the weight of the root cgroup.
7075 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
7077 mutex_lock(&shares_mutex);
7078 if (tg->shares == shares)
7081 tg->shares = shares;
7082 for_each_possible_cpu(i) {
7083 struct rq *rq = cpu_rq(i);
7084 struct sched_entity *se;
7087 /* Propagate contribution to hierarchy */
7088 raw_spin_lock_irqsave(&rq->lock, flags);
7090 /* Possible calls to update_curr() need rq clock */
7091 update_rq_clock(rq);
7092 for_each_sched_entity(se)
7093 update_cfs_shares(group_cfs_rq(se));
7094 raw_spin_unlock_irqrestore(&rq->lock, flags);
7098 mutex_unlock(&shares_mutex);
7101 #else /* CONFIG_FAIR_GROUP_SCHED */
7103 void free_fair_sched_group(struct task_group *tg) { }
7105 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7110 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
7112 #endif /* CONFIG_FAIR_GROUP_SCHED */
7115 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
7117 struct sched_entity *se = &task->se;
7118 unsigned int rr_interval = 0;
7121 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
7124 if (rq->cfs.load.weight)
7125 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
7131 * All the scheduling class methods:
7133 const struct sched_class fair_sched_class = {
7134 .next = &idle_sched_class,
7135 .enqueue_task = enqueue_task_fair,
7136 .dequeue_task = dequeue_task_fair,
7137 .yield_task = yield_task_fair,
7138 .yield_to_task = yield_to_task_fair,
7140 .check_preempt_curr = check_preempt_wakeup,
7142 .pick_next_task = pick_next_task_fair,
7143 .put_prev_task = put_prev_task_fair,
7146 .select_task_rq = select_task_rq_fair,
7147 .migrate_task_rq = migrate_task_rq_fair,
7149 .rq_online = rq_online_fair,
7150 .rq_offline = rq_offline_fair,
7152 .task_waking = task_waking_fair,
7155 .set_curr_task = set_curr_task_fair,
7156 .task_tick = task_tick_fair,
7157 .task_fork = task_fork_fair,
7159 .prio_changed = prio_changed_fair,
7160 .switched_from = switched_from_fair,
7161 .switched_to = switched_to_fair,
7163 .get_rr_interval = get_rr_interval_fair,
7165 #ifdef CONFIG_FAIR_GROUP_SCHED
7166 .task_move_group = task_move_group_fair,
7170 #ifdef CONFIG_SCHED_DEBUG
7171 void print_cfs_stats(struct seq_file *m, int cpu)
7173 struct cfs_rq *cfs_rq;
7176 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
7177 print_cfs_rq(m, cpu, cfs_rq);
7182 __init void init_sched_fair_class(void)
7185 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7187 #ifdef CONFIG_NO_HZ_COMMON
7188 nohz.next_balance = jiffies;
7189 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7190 cpu_notifier(sched_ilb_notifier, 0);